Glossary
Every hard term the posts define, in one place: 494 terms, each with one plain definition and the posts where it appears. In a post, the same definitions open when you hover over or tap a term with a dotted underline.
494 terms
#
- .NET
Microsoft's software platform for building applications, most often in the C# language.
- 16-bit
A number format that uses 16 bits, or two bytes, per value: smaller than the usual 32-bit format, at the cost of some precision.
- 401 Unauthorized
The HTTP status code meaning the server refused the request because it did not carry valid credentials.
- 403 Forbidden
The HTTP status code for Forbidden: the server knows who you are, but that identity is not allowed to do this.
- 429 Too Many Requests
An HTTP status code meaning too many requests: the caller hit a rate limit or quota and should wait before retrying.
A
- A2A (Agent2Agent)
Agent2Agent: an open protocol that lets AI agents built by different teams or vendors discover each other and exchange tasks and messages over HTTP.
- Access Connector for Azure Databricks
An Azure resource that holds a managed identity, an Azure-run identity with no stored secret, which Databricks uses to reach Azure storage.
- access token
A short-lived digital pass, issued after someone signs in, that a program shows to a service to prove it may act for that person.
- ACL
Access control list: the list attached to a resource, such as a document, saying which users or groups may read or change it.
- ADK (Agent Development Kit)
Agent Development Kit: Google's open-source framework for building, testing and deploying AI agents and multi-agent systems in code.
- adversarial inputs
Inputs crafted on purpose to trick a model into a wrong or harmful output, often looking normal to a person.
- after-model callback
A function ADK runs straight after each model response, which can inspect the response and replace it before anything else sees it.
- agent card
A small public JSON file in which an A2A agent describes its name, skills, address and how callers must authenticate.
- Agent Gateway
The Agent Platform component that acts as the governed entry and exit point for agent traffic, allowing or refusing each call to a tool, model or agent by policy.
- Agent Identity
A Google Cloud feature that gives each deployed agent its own cryptographic identity, based on the SPIFFE standard, so it can be granted roles and audited separately.
- Agent Registry
Google Cloud's central catalogue for registering, discovering and governing agents, MCP servers, tools and endpoints, which Agent Gateway can check calls against.
- Agent Runtime
Google Cloud's fully managed service for deploying, scaling and running agents in production, part of Agent Platform; formerly Vertex AI Agent Engine, a name some resources and SDKs still use.
- Agentic AI Foundation
A Linux Foundation body that hosts open agent standards, such as MCP, under neutral, multi-vendor governance.
- agents-cli
Google's command-line tool for creating, running, testing and deploying ADK agent projects from templates.
- aggregate
A summary figure, such as a total or average, computed over many rows instead of returning the rows themselves.
- AI Office
The body inside the European Commission that supervises general-purpose AI models and coordinates how the AI Act is applied across the EU.
- AI Omnibus
The Digital Omnibus on AI: an EU package of targeted amendments to the AI Act, proposed by the Commission in November 2025, which among other changes postpones the high-risk deadlines.
- Allowlist
A list of the only accounts, projects, destinations or actions that are permitted; anything not on it is refused.
- Amazon S3
Amazon Simple Storage Service, the object storage service on AWS, the counterpart of Cloud Storage.
- Apache 2.0
A permissive open-source licence that allows commercial use, changes and redistribution, as long as you keep the licence and copyright notices.
- Apache Flink
An open source engine for processing streams of data continuously as events arrive, rather than in scheduled batches.
- Apache Hive
Open source software that lets people query files in distributed storage with SQL, and whose metastore many engines use as a table catalogue.
- Apache Iceberg
An open table format, similar to Delta Lake, that stores a table as data files plus metadata in object storage, so engines such as BigQuery and Spark can share it.
- Apache Spark
An open source engine that splits large data processing jobs across many machines in a cluster working in parallel.
- API (application programming interface)
Application programming interface: a defined way for one program to ask another for data or actions, such as a web address that returns prices.
- API key
A secret string an app sends with each request to a paid service so the service knows whose account to use and bill.
- Apigee
Google Cloud's API management platform, which sits in front of APIs to handle security, rate limits, analytics and routing.
- App Check
A Firebase service that checks requests really come from your genuine app or site, so others cannot use your backend resources or quota.
- Application Load Balancer
Google Cloud's load balancer for HTTP and HTTPS traffic, which can terminate TLS and route each request by its host name or path.
- ARIMA
AutoRegressive Integrated Moving Average: a classic statistical method that forecasts a series from its own recent values and its recent forecast errors.
- attribute mappings
Rules that copy details from the external identity provider's sign-in token, such as user ID or group, into the attributes Google Cloud uses for access decisions.
- attribution graph
A map, built by Anthropic's circuit tracing method, of which internal features passed information to which others on one prompt, tracing how an input becomes an output.
- auto-regressive
Generating a sequence one piece at a time, feeding each new piece back in as input before producing the next.
- autocorrelation
When a value is similar to the values just before it in time, as this week's orders usually are to last week's.
- Autoscaling
Automatically adding or removing computing capacity, such as cluster workers, as demand rises and falls, usually between a minimum and maximum that you set.
- Availability Zone
AWS's name for one or more isolated data centres inside a Region, with separate power and networking, much like a Google Cloud zone.
- AWS Glue
AWS's managed data integration service, whose Glue Data Catalog records tables, their schemas and where their files sit in Amazon S3.
B
- B-tree
A sorted, branching structure that lets a database find a value in a few steps instead of scanning every row.
- backfill
A one-off run that processes past periods of data, for example to fill a gap or apply a corrected rule to history.
- baremetal
A physical server used directly, with no hypervisor or virtual machine layer between the hardware and the operating system.
- base model
A model straight out of pretraining, before any tuning to follow instructions or chat; it simply continues the text it is given.
- base64
A way of writing any binary data using only 64 printable characters, so it can travel safely inside text such as a web form.
- batch inference
Running a model over many stored records in one go, rather than answering live requests one at a time.
- batch job
Work submitted to run in the background over many inputs at once, rather than answering one request at a time while someone waits.
- bearer token
An access token that grants access to whoever presents it, like a ticket, so it must be kept secret and short-lived.
- benchmark
A fixed public test set, such as MMLU or GPQA, used to score and compare models on the same questions.
- BIG-Bench
A collaborative benchmark of over 200 varied tasks, from arithmetic to translation puzzles, contributed by many researchers to probe the abilities of language models.
- BigQuery
Google Cloud's serverless data warehouse, where companies store large tables and query them with SQL.
- BigQuery Agent Analytics
An ADK plugin that streams agent events, such as prompts, tool calls and responses, into a BigQuery table for analysis.
- Bigtable
Google Cloud's managed wide-column NoSQL database, built for very large tables and fast reads and writes of single rows by key.
- binaries
Compiled program files that a computer can run directly, as opposed to the human-readable source code they were built from.
- biometric data
Personal data from measuring a person's physical or behavioural features, such as their face, fingerprints or voice, that can identify or categorise them.
- BLEU
A translation score from 0 to 100 that measures how many word sequences in the machine's output also appear in human reference translations; higher is better.
- broadcast address
The last address in a range, traditionally used to send one message to every device on that subnet at once.
- browser console
The browser's developer tool where anyone can type and run JavaScript against the page they have open.
C
- calibrated
A model is calibrated when its scores match reality: of all the cases it scores about 0.7, about 70% turn out to be positive.
- callback
A function ADK runs automatically at a fixed point in a turn, such as just before a tool call, which can inspect or block that step.
- canary
A release that sends a small share of real traffic to the new version first, and widens it only if nothing goes wrong.
- catalog
In Unity Catalog, the top level container of data, holding schemas, which in turn hold tables and views.
- certificate authority (CA)
An organisation or service that signs certificates; a device accepts a certificate only if it was signed by an authority on that device's trusted list.
- change data feed
A Delta table option that records row-level inserts, updates and deletes between versions, so downstream jobs can process only what changed.
- chargeback
Billing each customer or team for exactly the resources they used.
- checkpoint
A Parquet file Delta Lake writes now and then to summarise the table's log so far, so readers need not replay every JSON commit.
- chunking
Splitting long documents into smaller passages so each one can be indexed and retrieved on its own.
- CI/CD (continuous integration and delivery)
Continuous integration and continuous delivery: an automated pipeline that builds, tests and often deploys code every time a change is pushed, so nobody applies changes by hand.
- circuit breaker
A switch in code that stops calling a failing service for a while after repeated errors, then lets a few trial calls through to test recovery.
- classic compute plane
Where classic Databricks clusters run, as virtual machines inside your own Google Cloud project and network.
- classifier
A model that sorts an input into categories, such as harmful or harmless, usually by giving a score and comparing it with a cutoff.
- CLI
Command-line interface: a program you control by typing commands in a terminal rather than clicking in a window.
- Cloud Firestore
Google's managed document database, part of Firebase, which can push changes to connected apps in real time.
- Cloud Interconnect
A dedicated private physical link between an on-premises data centre and Google's network.
- Cloud KMS
Google Cloud Key Management Service: it stores encryption keys and controls which identities may use them to encrypt or decrypt data.
- Cloud Logging
Google Cloud's managed service that collects, stores and lets you search log lines from your applications.
- Cloud NAT
Google Cloud's managed network address translation, which lets machines with only private IP addresses make outbound connections to the internet.
- Cloud Run
Google Cloud's managed service for running containers: you supply the container, and it handles servers and scaling, down to zero when idle.
- Cloud Scheduler
Google Cloud's managed cron service, which sends a message or calls a URL on a fixed schedule.
- Cloud Storage
Google Cloud's object storage service; files are kept in containers called buckets.
- Cloud Trace
Google Cloud's service for collecting and viewing traces, the timed, step-by-step record of each request.
- CMEK (customer-managed encryption keys)
Customer-managed encryption keys: data is encrypted with keys you create and control in your own key service, such as Cloud KMS, so you can revoke access.
- cold start
The delay while a stopped service or machine starts up before it can answer its first request.
- Committed-use
Committed-use discount: a lower price in exchange for committing to a minimum spend or usage, usually for one or three years.
- Compaction
Shortening a long conversation history, typically by replacing older turns with a short model-written summary, so each model call sends fewer tokens.
- Compute Engine
Google Cloud's service for creating and running virtual machines.
- concurrency control
How a database keeps data correct when many users read and change the same records at the same moment.
- conditional edge
A link in a graph whose destination is chosen while the program runs, by a function that looks at the current state.
- consent screen
The page where a user agrees to let an app act on their behalf with a specific service, before any access token is issued.
- Container
A lightweight package of an application with the code, libraries and settings it needs, so it runs the same way on a laptop, a server or a cloud service.
- container runtime
The software on a machine that actually starts, runs and stops containers from their images, such as Docker or containerd.
- Context caching
Reusing the already processed opening part of a model request when later requests start with the same text, such as fixed instructions, which cuts cost and latency.
- context window
The maximum amount of text, measured in tokens, that a model can read in one call, including instructions, history and tool results.
- Context-Aware Access
A Google Cloud control that allows or refuses a request based on conditions about it, such as the device, network or certificate it comes from.
- Control plane
The part of a system that manages and coordinates the work, such as the web interface, APIs and scheduling, as opposed to the machines where the work itself runs.
- CopilotKit
An open-source framework for building in-app AI assistant interfaces, which created and maintains the AG-UI protocol.
- CORS
Cross-Origin Resource Sharing: the browser rules that decide whether a page on one domain may read responses from a server on another domain.
- COTS
Commercial off-the-shelf: ready-made software bought from a vendor rather than built in-house.
- covariate shift
A change in the distribution of a model's inputs while the relationship between inputs and the answer stays the same.
- CPM
Cost per mille: advertising spend for every thousand impressions, so spend divided by impressions, times 1,000.
- CrewAI
An open-source Python framework for building teams of AI agents that each play a defined role.
- CRM
Customer relationship management: software that keeps track of a company's customers, contacts, sales deals and service cases.
- cron
The classic Unix scheduler, and its compact five-field notation for times such as every day at 02:00.
- cross-entropy loss
A score of how surprised a language model is by the real next token in unseen text, averaged over many tokens; lower means better predictions.
- cross-layer transcoder
A readable stand-in network trained to imitate a model's MLP layers using sparse features, where each feature reads from one layer and writes to all later ones.
- cryptographic provenance
A digitally signed record attached to a file, as in the C2PA standard, showing where the content came from and how it was edited.
- CSAT (customer satisfaction score)
Customer satisfaction score: the share of customers who rate an interaction as good in a short survey sent straight after it.
- CSV
Comma-separated values: a plain text file where each line is a row and commas separate the columns.
- CTR (click-through rate)
Click-through rate: the share of times an ad was shown, called impressions, that ended in someone clicking it, so clicks divided by impressions.
- custom-mode
A custom-mode VPC network starts with no subnets; you create each subnet and choose its region and IP range yourself.
- customer-managed VPC
A network you create and control in your own Google Cloud project, which Databricks then uses for a workspace's classic clusters.
D
- Data pipeline
A chain of automated, often scheduled, steps that copies data out of its sources, cleans and reshapes it, and loads it where it will be used.
- data poisoning
An attack that slips misleading or malicious examples into training data so the resulting model behaves wrongly.
- data residency
The requirement that data is stored, and sometimes processed, only in a chosen country or region.
- Databricks
A data and AI platform, built around Apache Spark, where companies store tables, run analytics and train models; it runs on Google Cloud, AWS and Azure.
- Databricks AI Search
Databricks' managed vector search service, formerly Vector Search: it finds stored documents or records whose meaning is closest to a question, not just those sharing its words.
- Databricks Runtime
The versioned bundle of Apache Spark and libraries that runs on Databricks compute; each version number is a release of it.
- Databricks service principal
A Databricks identity for an automated tool, job or program rather than a person, with its own permissions and credentials.
- DataFrames
A pandas table held in memory, with named columns and rows, that Python code can filter, sum and plot.
- dbt Cloud
The hosted service for dbt, a tool that builds, tests and documents data transformations written in SQL.
- DBU (Databricks Unit)
Databricks Unit: the normalised unit of processing Databricks measures and bills by; each product consumes DBUs at its own rate, and you pay a price per DBU.
- DDOS
Distributed denial of service: an attack where many machines flood a service with traffic at once so genuine users cannot reach it.
- decoder
The half of a translation model that writes the output one token at a time, looking at what it has written so far and at the encoder's vectors.
- default gateway
The address a machine sends traffic to when the destination lies outside its own subnet.
- Delta Lake
An open table format that adds a transaction log to Parquet files in cloud storage, giving reliable updates, deletes and version history.
- Delta table
A table stored as Parquet data files plus a transaction log in object storage, using the open Delta Lake format.
- deprecation warning
A notice from an API that a feature still works for now but is being retired and will stop working later.
- differential privacy
A mathematical guarantee that a published statistic changes very little whether or not any one person's data is included, usually achieved by adding carefully sized random noise.
- distributed system
A system whose data and work are spread across many machines that coordinate over a network.
- DML and DDL
Data manipulation language (INSERT, UPDATE, DELETE) changes the rows in tables; data definition language (CREATE, ALTER, DROP) changes the tables themselves.
- DNS
The Domain Name System, which turns names such as a web address into the IP addresses that machines actually connect to.
- DNS peering
A Cloud DNS setting that answers lookups for chosen domain names using another VPC network's DNS settings, so private names resolve to private addresses.
- Docker container
A packaged app with its own files and libraries that runs as an isolated process, but still shares the host machine's operating system kernel.
- Docstring
The text written just below a Python function's first line that says what it does, its arguments and what it returns; ADK shows it to the model.
- Document AI
Google Cloud's service for extracting text, layout, tables and fields from documents such as scanned PDFs, forms and invoices.
- DOM
Document Object Model: the browser's live, in-memory tree of every element on the page, which JavaScript reads and changes.
- dot product
Multiply two vectors number by number and add up the results; the bigger the total, the more the two vectors point the same way.
- DPoP (Demonstrating Proof of Possession)
Demonstrating Proof of Possession: an OAuth extension in which the client signs each request with its own private key, so a stolen token is useless without that key.
- Drift
A gradual change over time in the data a model or system sees, or in its behaviour, so patterns it learned or was tested on stop matching.
E
- eager execution
Running each operation immediately as the code reaches it, which is easy to debug, instead of first compiling the whole program into one optimised graph.
- early stopping
Stopping training once the score on held-back data stops improving, so the model does not keep memorising the training rows.
- Egress
Network traffic leaving a system, network or zone for somewhere outside it; cloud providers often charge for it, and policies can restrict where it may go.
- Embedding
A list of numbers that captures the meaning of a piece of text or other content, so that similar meanings end up close together and can be found by search.
- encoder
The half of a translation model that reads the whole input sentence and turns it into vectors.
- Endpoint
The network address, usually a URL, where a service such as a deployed model accepts requests and returns results.
- Environment variable
A named setting a running program reads from its operating system, often used to pass secrets such as API keys or paths without writing them into code.
- Envoy
An open-source network proxy, now a Cloud Native Computing Foundation project, that many gateways and load balancers are built on to route, secure and observe traffic.
- ERP
Enterprise resource planning: the core business system that runs orders, stock, finance and purchasing, such as SAP or Oracle.
- EU Member States
The 27 countries of the European Union, whose governments must approve EU laws and amendments alongside the European Parliament.
- European Commission
The European Commission: the EU's executive body, which proposes laws and publishes guidance on how they should be applied.
- Eval
Short for evaluation: a repeatable test that runs fixed questions through an AI system and scores its answers and behaviour against agreed expectations.
- Eval gate
An automated check that runs a fixed set of test questions and blocks a change from release if the scores fall below an agreed bar.
- event-driven
Built so that steps start in response to events, such as a message arriving or a previous step finishing, rather than following one fixed script.
- exact match
A scoring rule that gives full marks only if the model's output matches the reference answer character for character, and zero otherwise, with no partial credit.
- exception
An error a Python program raises when something goes wrong; if nothing catches it, the running code stops.
- exfiltrate
To copy data out of a system without permission, for example by sending it to a server the attacker controls.
- External location
A Unity Catalog object that pairs a cloud storage path with a storage credential, so access to files under that path can be granted and audited.
F
- f-string
A Python string written with an f prefix that inserts variable values directly into the text, which is unsafe for building SQL from user input.
- failover
Switching automatically to a backup, such as another model, provider or server, when the first one fails or is unavailable.
- faithful (chain of thought)
Describes a chain of thought that accurately reflects the factors that actually produced the model's answer, including any shortcut or hint it relied on.
- FastAPI
A popular open-source Python framework for building web APIs quickly, with request checking based on Python type hints.
- FCA
Financial Conduct Authority: the UK regulator for financial services firms and the promotions they publish.
- federating
Querying data where it already lives, from another system, instead of copying it into one place first.
- few-shot prompting
Showing a model a few worked examples inside the prompt so it copies the pattern, with no change to the model's weights.
- fine-tuning
Further training of an existing model on a smaller, task-specific dataset, which changes its weights and so its behaviour.
- FIPS 140
A US government standard for testing and certifying that the modules performing encryption do it correctly and securely.
- Firebase
Google's platform for building web and mobile apps, providing hosting, sign-in and databases as managed services.
- firewall rule
A rule on a VPC network that allows or blocks traffic to or from virtual machines, by address range, port and protocol.
- FLOPs
Floating point operations: individual multiplications or additions on decimal numbers. Training compute is counted as the total number of them, often written like 5.76 x 10^23.
- FOIA
Freedom of Information Act: the US law that lets the public request records held by federal agencies.
- foreign catalog
A Unity Catalog catalog that mirrors the databases and tables of an external system, so Databricks users can query them by name.
- frontend
The part of an application the user sees and interacts with, such as a web page or mobile app.
G
- garbage collection
A runtime feature that automatically frees memory a program is no longer using.
- gcloud
The Google Cloud command line tool, used to create, inspect and manage resources from a terminal or script.
- GDPR
General Data Protection Regulation: the EU law on how organisations collect, use, store and protect personal data.
- Gemini Enterprise
Google Cloud's workplace AI product, combining an AI assistant, company-wide search and agents, where employees chat with Gemini and use agents their organisation publishes.
- Gemini Enterprise Agent Platform
Gemini Enterprise Agent Platform: Google Cloud's unified platform for building, deploying, governing and optimising AI agents and models, the evolution of Vertex AI.
- Gemini Flash
Gemini Flash: the faster, cheaper family of Gemini models, suited to simple, high-volume steps.
- Gemma 2
Google's family of open models, with published weights, released in 2024 in 2B, 9B and 27B parameter sizes.
- General availability (GA)
General availability: the launch stage, after Preview, at which a product is fully supported for production use, with its published service level agreements.
- Genie
Databricks' natural language experience that lets business users ask questions of governed data and answers them by generating SQL over Unity Catalog tables.
- Git
A version control system that records every change to a set of files, who made it and when, so changes can be reviewed and undone.
- GitHub Actions
GitHub's built-in service for running automated workflows, such as tests and deployments, whenever code changes.
- GKE (Google Kubernetes Engine)
Google Kubernetes Engine: Google Cloud's managed service for running containers on Kubernetes clusters.
- Golden set
A fixed set of real test questions with checked, correct answers, rerun after every change so the system is scored the same way and regressions show up.
- Google Gen AI Python SDK
Google's Gen AI software development kit for Python: a library of ready-made functions for calling Gemini from code.
- GPQA
A benchmark of very hard, graduate level multiple choice science questions written by experts and designed to be hard to answer even with a web search.
- GPU
Graphics processing units: chips with thousands of small cores that do many simple calculations at the same time, which is why they suit training AI models.
- gradients
The numbers training computes to say which way, and by how much, each trainable weight should move to reduce the error.
- grants
Permissions given to a user or group on an object, such as the right to read one table.
- graph workflow
An agent built as a graph: each node is a step or an agent, and each edge says which step runs after which.
- grounding
Making a model base its answer on supplied sources, such as retrieved documents, rather than on memory alone.
- gRPC
An open-source framework, created at Google, for one program to call functions in another over the network using compact binary messages.
- GSM8K
A public test set of about 8,500 grade school maths word problems, each needing a few steps of arithmetic, widely used to measure reasoning over several steps in language models.
- guardrails
Checks around an AI system that block or correct unsafe, off-topic or policy-breaking inputs and outputs.
- gVisor
An open-source Google sandbox that sits between a container and the host operating system, intercepting system calls so the code cannot touch the real kernel directly.
H
- Hadoop
An older open-source framework for storing and processing very large datasets across a cluster of ordinary servers.
- hallucinates
Produces something that sounds plausible but is false or not supported by its input, stated as confidently as a fact.
- hard-coded
Written directly into the code as a fixed value, so it does not change unless someone edits the code.
- harness
The code around a model that runs it: it builds each prompt, calls the model, runs the tools it asks for and decides when to stop.
- headers
Named fields sent with a network request, separate from its body, often used to carry credentials and other details about the call.
- held-out eval set
Test questions kept apart from everything used to build or tune the system, so the score reflects how it handles questions it has not seen.
- HIPAA
The US Health Insurance Portability and Accountability Act, which sets rules for protecting patients' health information.
- Hive metastore
The older table catalogue, from Apache Hive, that each Databricks workspace kept for itself before Unity Catalog moved governance to the account level.
- horizontal scalability
Growing capacity by adding more machines, rather than buying one bigger machine.
- HTTP
Hypertext Transfer Protocol: the request and response language browsers and web servers use; HTTPS is the same protocol wrapped in encryption.
- hybrid search
Search that combines keyword matching with embedding similarity and merges the two ranked lists into one.
- hypervisor
The software layer that splits one physical server into several virtual machines and keeps each one isolated from the others.
I
- IAM (Identity and Access Management)
Identity and Access Management: Google Cloud's system for granting roles, and so permissions, to people, service accounts and agents on specific resources.
- IAM allow policy
A Google Cloud rule set attached to a resource that lists which identities hold which IAM roles on it, and so what they may do.
- IAM role
A named set of Google Cloud permissions that you grant to a person, group, service account or agent on a resource.
- ICT
Information and communications technology: the computers, networks, software and services an organisation runs on.
- ID token
A short-lived signed token, usually a JWT, that proves who the caller is to the service receiving the request.
- IDE
Integrated development environment: the code editor a developer works in, such as VS Code.
- Idempotency key
A unique value sent with a request so that, if the same request arrives twice, the service performs the action only once.
- identity provider
The system that signs people in and knows which groups they belong to, such as Microsoft Entra ID, Okta or Google Identity.
- IETF
The Internet Engineering Task Force, the open standards body that develops core internet protocols such as HTTP and TLS.
- implicit caching
Automatic reuse of a prompt's repeated opening text that the platform has seen recently, billed at a lower rate without you setting anything up.
- in-context learning
A model picking up a pattern or task from examples inside the prompt itself, with no change to its weights.
- indemnification
A promise in a contract to cover the other party's legal costs and damages if a specific kind of claim is made against them.
- ingress rules
VPC Service Controls rules that allow specific identities or networks outside a perimeter to reach the protected services inside it.
- init script
A shell script that a Databricks cluster runs on every node as it starts, for example to install software or set configuration.
- inpainting
Editing only a chosen area of an image, such as the background, while the rest of the picture, such as the product, stays as it was.
- input schema
A machine-readable description of the arguments a tool accepts, with each field's name and type, so the model knows how to call it.
- instance profiles
An AWS wrapper around an IAM role that lets programs on a virtual machine use that role's permissions without any stored keys.
- Interactions API
A Gemini API interface for calling models and agents, such as the Deep Research agent; long tasks can run in the background while the caller polls for the result.
- internal load balancer
A Google Cloud load balancer with a private IP address that spreads traffic from inside your network across the machines behind it.
- interpretability
Research that looks inside a model's internal numbers and pathways, such as features and circuits, to explain how it actually computes an answer.
- invocation
In ADK, everything that happens to answer one user message, from receiving it to the final reply.
- IPv4
Internet Protocol version 4: the original internet addressing scheme, written as four numbers from 0 to 255 separated by dots, with about 4.3 billion possible addresses.
- IPv6
Internet Protocol version 6: the newer addressing scheme with vastly more addresses than IPv4, written as eight groups of hexadecimal digits separated by colons.
- ISO 8601
The international standard way to write dates and times, such as 2026-10-20T14:05:00Z, so there is no doubt about format or time zone.
- ISP
Internet service provider: the company, such as BT or Virgin Media, that connects your home or office to the internet.
J
- jailbreak
A prompt written to trick a model into ignoring its safety rules, often through role play, disguise or a made-up scenario.
- JDBC and ODBC
Standard database driver interfaces, JDBC for Java and ODBC for most other languages, that BI tools and programs use to connect to a database and run SQL.
- Jensen-Shannon divergence
A number between 0 and 1 for how different two distributions are; 0 means identical. Model Monitoring uses it to compare live feature values with a baseline.
- join keys
The columns used to match rows in one table with rows in another, such as brand, region and week.
- JSON
JavaScript Object Notation, a plain text format that stores data as names and values, readable by people and programs.
- JSON Pointer
A short path string, standardised in RFC 6901, that picks one value out of a JSON document, such as /results/0/title.
- JSON-RPC
A lightweight protocol for asking another program to run a named method, with the request and the reply both written as JSON.
- JWT (JSON Web Token)
JSON Web Token: a compact, signed set of claims, such as who the user is and when access expires, that anyone holding it can read and the receiver can verify.
K
- k-means
A clustering method that places a chosen number of centre points, assigns each item to its nearest centre, then moves each centre to the middle of its group and repeats.
- key-value
Data stored as pairs of a name and its value, such as currency set to EUR, looked up by the name.
- knowledge graphs
Databases that store facts as entities and the relationships between them, such as retailer, has contract, contract version, so software can follow the links.
- KPI
Key performance indicator: a headline number a business tracks to judge how something is doing, such as total spend.
- Kubernetes cluster
A group of machines managed by Kubernetes, open-source software that starts, restarts and scales containers across them.
- Kubernetes Job
A Kubernetes task that starts one or more containers, runs them until the work finishes, then stops.
L
- L-infinity distance
For categories, the largest difference between the share of any one category in the two datasets, for example 30% of homes rural in training and 55% now gives 0.25.
- Lakeflow
Databricks' family of data engineering tools for ingesting data, building data pipelines and orchestrating jobs.
- lakehouse
A data platform that keeps tables as open files in low-cost object storage, then adds warehouse features such as transactions, governance and fast SQL.
- Lakehouse Federation
A Databricks feature that gives governed, read-only access to tables in outside systems, such as BigQuery, through Unity Catalog foreign catalogs, without copying the data.
- landing-zone
A landing zone is the pre-built foundation of a cloud estate: hierarchy, networks, policies and access set up before workloads move in.
- LangGraph
An open-source framework from the LangChain team for building agents as graphs of steps with shared state.
- LangSmith
LangChain's commercial platform for tracing, evaluating and deploying agents built with LangChain and LangGraph.
- latency
The delay between sending a request and getting the response back, usually measured in milliseconds.
- leased line
A private, dedicated connection rented from a telecoms provider between two sites, offering steady bandwidth that is not shared with public internet traffic.
- lineage
A record of where data came from and which tables, jobs, notebooks and dashboards it flows into.
- lint
An automatic check that reads source code without running it and flags style problems and likely mistakes.
- LiteLLM
An open-source Python library that offers one interface for calling models from many providers, such as OpenAI, Anthropic or a vLLM server.
- LLM
Large language model: a model trained on huge amounts of text to predict the next token, such as Gemini.
- local SSD
A solid state drive physically attached to the machine, much faster than network storage, whose data is lost when the machine is stopped.
- logistic regression
A simple model that multiplies each input by a learned weight, adds them up and squashes the total into a probability between 0 and 1.
- Looker
Google Cloud's business intelligence platform for dashboards and reports built on one shared, governed definition of the business metrics.
M
- managed identity
An Azure identity for a resource or program whose credentials Azure creates and rotates itself, so no secret is stored in code.
- managed tables
Unity Catalog tables whose files Databricks stores and looks after in governed storage, and deletes when the table is dropped.
- Markdown
A simple plain-text format that uses symbols such as # for headings and * for lists, readable as it is and easy to turn into a web page.
- market surveillance authority
The national regulator in each EU country that checks products on the market comply with the law and can demand evidence or withdrawals.
- MCP (Model Context Protocol)
Model Context Protocol: an open standard for exposing tools and data to AI agents in one consistent way.
- Medallion architecture
A common way to organise lakehouse tables in stages: raw data lands in bronze, is cleaned in silver and is shaped into business-ready tables in gold.
- Memory Bank
Gemini Enterprise Agent Platform's managed long-term memory: it generates memories about a user from conversations and recalls them in later sessions, so an agent can personalise answers.
- Metadata
Data that describes other data, such as who created a file and when, kept alongside the content rather than in it.
- metadata server
A service built into every Compute Engine virtual machine that answers local requests for instance details and short-lived access tokens for its service account.
- Metastore
A catalogue of table metadata; in Unity Catalog, the top-level container for a region that holds catalogs, schemas, tables and their permissions, which workspaces attach to.
- Metric views
Unity Catalog objects that store a business metric's agreed formula and dimensions once, so every query and tool uses the same definition.
- metrics
The rule used to turn a model's answers into a score, such as exact match, partial credit per token, or a score based on probabilities.
- Microsoft Entra ID
Microsoft's cloud identity service, formerly Azure Active Directory, which signs employees in and manages their accounts and groups.
- Middleware
Software that sits between applications and the operating system or other systems, such as a web server or message broker, providing shared services.
- MIT licence
A short, permissive open-source licence: anyone may use, copy, change and sell the code, as long as they keep the copyright notice.
- MLflow
An open-source platform, created by Databricks, for tracking machine learning and AI experiments, tracing agents, evaluating quality and managing models.
- MLP layer
Multi-layer perceptron: the part of each Transformer layer that transforms each token on its own after attention has mixed in information from other tokens.
- MMLU
Massive Multitask Language Understanding: a benchmark of multiple choice questions across 57 subjects, from law to physics, used to compare language models' knowledge and reasoning.
- model alias
A model name, such as one ending in latest, that the provider can point at a newer model without you changing any code.
- Model Armor
A Google Cloud service that screens prompts and model responses for risks such as prompt injection, jailbreaks, sensitive data, harmful content and malicious URLs.
- Model card
A document a model provider publishes describing what a model is for, how it was trained and evaluated, its known limits and safety testing.
- Model Garden
Google Cloud's catalogue of Google, partner and open models, from which you can call managed APIs or deploy open models to your own endpoints.
- Model Serving
Databricks' managed service for deploying, governing and querying machine learning and generative AI models behind REST API endpoints, for real-time and batch use.
- model-agnostic
Able to work with models from different providers, not only the one it was designed around.
- mTLS (mutual TLS)
Mutual TLS: an encrypted connection where both sides show a certificate, so each knows who the other is.
N
- Network peering
Joining two private networks, such as two VPC networks, so their internal IP addresses can reach each other directly without crossing the internet.
- network tab
The part of the browser's developer tools that lists every request a page makes and shows the data each one returns.
- network tag
A label on a VM that firewall rules and routes can target, so a rule applies only to machines carrying that tag.
- next hop
Where a route sends matching traffic next, such as an internet gateway, a VM or a VPN tunnel.
- Notebook
An interactive document, such as a Jupyter or Databricks notebook, that mixes runnable code cells with their results and notes, used to explore data and build jobs.
O
- O*NET
The Occupational Information Network: a US Department of Labor database describing around 20,000 work tasks and the occupations that perform them.
- OAuth 2.0
The standard protocol for letting an application act on a user's behalf with limited, revocable permission, using short-lived access tokens instead of passwords.
- observability
Being able to see what a running system did and why, from the traces, logs and metrics it records.
- on-call rota
The schedule of who is responsible, at any hour, for responding when a production system raises an alert.
- On-premises
Running in a company's own buildings or data centres, on hardware it owns and operates, rather than in a public cloud.
- ontologies
Formal, agreed definitions of the types of things in a domain and how they relate, which give a knowledge graph its vocabulary and rules.
- open-weight model
A model whose trained weights are published for download, so you can run it on your own servers or cloud instead of calling the maker's API.
- OpenAPI
A standard, machine-readable way to describe a REST API: its paths, parameters and request and response formats, so tools can call it.
- OpenID Connect (OIDC)
OpenID Connect: an identity standard built on OAuth 2.0 in which the identity provider issues a signed ID token proving who a user or service is.
- OpenSharing
Databricks' open protocol, formerly called Delta Sharing, for sharing live tables and other assets with other organisations and platforms without copying the data.
- OpenTelemetry
An open-source standard and toolkit for producing traces, metrics and logs in one format that many monitoring tools can read.
- orchestration
Software that automatically starts, stops, scales and places workloads on servers according to demand.
- Orchestrator
A coordinating program or agent that splits work into tasks, triggers each one in the right order on the right system or agent, and combines the results.
- organisation policy
A Google Cloud rule, set on an organisation, folder or project, that restricts how resources below it may be configured.
- OTIF (on time in full)
On time in full: the share of orders delivered on the agreed date with every item and quantity complete, a headline supply chain service measure.
- over-refusal
When a model or safety filter declines a harmless request because it wrongly looks risky, such as refusing a question about mustard because mustard gas exists.
P
- p95 (95th percentile)
The 95th percentile: the value that 95 in every 100 results fall at or below, so for response times only the slowest 5 per cent take longer.
- pandas
A popular Python library for loading, reshaping and summarising tables of data.
- Parameterised query
A database query where user values are sent separately from the SQL text as parameters, so they are always treated as data and can never change the command.
- Parquet
An open file format that stores data column by column, compressed, so analytical queries read only the columns they need.
- PCI-DSS
The Payment Card Industry Data Security Standard, which any company that stores, processes or transmits card data must meet.
- PECR
The Privacy and Electronic Communications Regulations: the UK rules on electronic marketing such as emails and texts, and on cookies, which sit alongside UK GDPR.
- PEP 604
The Python Enhancement Proposal that lets you write a union of types with a vertical bar, such as str | None, supported from Python 3.10.
- persistent disk
Network-attached block storage for Compute Engine, whose data lives independently of the virtual machine it is attached to.
- Photon
The Databricks native vectorised query engine, written in C++, which runs SQL and DataFrame work faster than standard Spark.
- PKCE
Proof Key for Code Exchange: an OAuth add-on in which the app proves it started the sign-in, designed for apps that cannot keep a secret.
- post-training
The training stages after pretraining, such as tuning on example answers and reinforcement learning, that turn a base model into a helpful assistant.
- power law
A relationship where one quantity changes as a fixed power of another, so it appears as a straight line when both are plotted on logarithmic scales.
- PPO
Proximal Policy Optimization: a reinforcement learning algorithm that improves a model in small, limited steps so each update cannot change its behaviour too much at once.
- Predictive optimisation
A Databricks service that automatically runs table maintenance, such as compaction and VACUUM, on Unity Catalog managed tables.
- preference model
A model trained on people's choices between pairs of answers, which then gives any new answer a score for how much people would prefer it.
- pretraining
The first and largest training stage, where a model learns to predict the next token across a huge collection of text.
- Preview
A Google Cloud launch stage before general availability: you can try the feature, but it may change and has limited support and no production service level agreement.
- private key
The secret half of a cryptographic key pair, kept only by its owner and used to prove identity or sign data.
- Private Service Connect
A Google Cloud feature for reaching a service through a private IP address inside your own VPC network, so traffic stays off the public internet.
- Pro
Gemini Pro: the larger, more capable family of Gemini models, slower and dearer than Flash, used where reasoning quality matters most.
- probe
A small, cheap classifier that reads a model's internal activations, the numbers inside it while it processes text, to spot a pattern such as harmful intent.
- prompt injection
An attack where text hidden in user input or a retrieved document tries to override the model's instructions, for example telling it to leak data.
- Protocol Buffers
Google's language-neutral format for describing structured data in a schema file and serialising it compactly between programs.
- Provisioned Throughput
A fixed-cost, fixed-term subscription that reserves model capacity for your own traffic, paid whether or not you use it, in return for guaranteed throughput.
- proxy
A server that forwards requests on behalf of other machines, often used as the single controlled exit point to the internet.
- Pub/Sub
Google Cloud's messaging service: one system publishes messages to a topic, and every subscribed system receives them.
- public key cryptography
Encryption that uses a pair of keys: a private key kept secret to sign or decrypt, and a public key anyone can use to verify or encrypt.
- Pull request
A proposed change to a code repository that others review and automated checks test before it is merged into the main branch.
- pure functions
Functions whose output depends only on their inputs and which change nothing outside themselves, so the same input always gives the same result and testing is easy.
- push notifications
Messages a server sends to a client's own web address when something changes, so the client does not have to keep asking.
- Pydantic
A Python library that defines data shapes as classes with typed fields and checks that incoming data matches them.
- PyPI
The Python Package Index, the public repository from which Python libraries are usually installed with pip.
- pytest
A widely used open-source Python testing framework: it finds functions named test_ and reports which pass and which fail.
Q
- Quantised (4-bit)
Stored with fewer bits per number, for example 4 instead of 16, which saves memory and speeds up a model at the cost of a little precision.
- quota
The amount of use, such as model calls per minute or per day, that a project is allowed before further requests are refused or billed.
R
- RAG
Retrieval-augmented generation: fetching relevant passages from your own documents and putting them in the prompt so the model answers from them.
- RAG Engine
A managed service in Gemini Enterprise Agent Platform that imports, chunks, embeds and indexes documents, then retrieves relevant passages so a model can ground its answers.
- ReAct
An agent pattern that loops: the model reasons about what to do, calls a tool, reads the result, and repeats until it can answer.
- React
A popular JavaScript library for building web interfaces out of reusable components.
- Reasoning model
A model that works through a problem in internal reasoning steps before giving its final answer, trading extra time and tokens for better answers on hard questions.
- red-teaming
Testing a system by deliberately attacking it the way an adversary would, such as trying to make an agent leak data or ignore its rules, to find weaknesses before real attackers do.
- refactor
Restructuring existing code to tidy or improve it without meaning to change what it does.
- refresh token
A longer-lived token the app keeps so it can get a fresh access token later without asking the user to sign in again.
- region
A specific geographic area, such as London or Iowa, where a cloud provider runs a cluster of data centres you can choose to place resources in.
- Regular expression
A compact text pattern, such as one matching any email address or run of digits, that code uses to find, check or replace text.
- reinforcement learning
A way of training where a model tries actions and is rewarded or penalised by results, learning to favour what earns rewards.
- rejection sampling
Generating several candidate answers and returning only the one a scoring model rates best; in this paper, the best of 16 by a harmlessness score.
- release candidate
A version believed to be ready for final release, published so people can test it before it is declared stable.
- release pipeline
The automated sequence of build, test and deploy steps that takes a code change from the repository to production.
- ReLU
A simple function used inside neural networks that passes positive numbers through unchanged and turns negative numbers into zero.
- reorder points
The stock level for a product at which a new order should be placed, set so the delivery arrives before the shelf runs empty.
- reranking
A second pass that reorders the retrieved passages by how well each one actually answers the question.
- reservations
A Compute Engine reservation holds VM capacity of a chosen machine type in one zone, so it is guaranteed to be there when you need it.
- residual stream
The running vector each token carries from layer to layer in a Transformer; every layer reads from it and adds its own result back.
- resolvers
DNS resolvers are the servers that look up names on a device's behalf and return the matching IP addresses.
- REST API
A common style of web API in which programs read and change resources by sending standard HTTP requests, such as GET or POST, to fixed web addresses.
- Resumability
An ADK app setting that saves a run's progress, so a run that is interrupted or paused for approval can continue later from the same point.
- retrieval
Searching a document store for the passages most relevant to a question, so they can be handed to the model with the question.
- reward hacking
When a model trained with rewards finds a loophole that earns the reward without doing the intended task, such as gaming a grader instead of answering correctly.
- RLHF
Reinforcement learning from human feedback: people compare pairs of answers, a model learns their preferences, and the assistant is trained to give answers that model scores highly.
- ROC-AUC
A score up to 1 for how well a classifier ranks positive cases above negative ones across every possible cutoff; 0.5 is no better than chance.
- Router (agent)
In the Acme example used across these series, the top-level agent that reads each question and hands it to the right specialist agent, such as marketing or supply.
- Row filters and column masks
Unity Catalog rules applied at query time: row filters hide rows a user may not see, and column masks hide or blank column values, depending on who runs the query.
- runbook
A written step-by-step guide that tells the on-call person what to check and do when a known problem happens.
- Runner
The ADK object that runs an agent: it takes the user's message, drives the model and tool calls, and saves events to the session.
S
- S&OP (sales and operations planning)
Sales and operations planning: the regular business process that balances demand forecasts against supply, production and stock plans.
- SaaS
Software as a service: software the vendor runs and maintains for you, used over the internet on a subscription.
- SAML
Security Assertion Markup Language: an older single sign-on standard in which the identity provider sends the application a signed XML statement about the user.
- sandboxes
Isolated environments where untrusted code can run without being able to reach or harm the rest of the system.
- SASE (Secure Access Service Edge)
Secure Access Service Edge: a cloud service between employees and the internet that combines network access with security checks such as web filtering and data loss prevention.
- scaffold
The starter project that agents-cli generates, with sample agent, test, eval and deployment files you then replace or extend.
- scale down to zero
Shut down every running copy of a service when no requests arrive, so you pay nothing while it is idle and it starts again on the next request.
- schema
In Unity Catalog, a named folder inside a catalog that groups related tables and views.
- schemas
The description of a table's structure: its column names, their data types and how the columns are arranged.
- SCIM
SCIM: System for Cross-domain Identity Management, a standard protocol an identity provider uses to create, update and remove users and groups in another system automatically.
- scratchpad
A private space where a model writes out its step-by-step reasoning before answering; researchers read it to see why the model acted as it did.
- SDK (software development kit)
Software development kit: ready-made libraries, tools and documentation a vendor publishes so your code can call its service without writing raw network requests.
- Secret Manager
Google Cloud's service for storing passwords, API keys and other secrets, with access controlled by IAM.
- self-signed certificate
A TLS certificate signed with its own key instead of by a certificate authority, so clients trust it only if they are told to explicitly.
- Semantic Kernel
Microsoft's open-source SDK for adding AI models, plugins and planning to C#, Python and Java applications.
- semantic layer
A shared set of business definitions, such as metric formulas and column meanings, that sits between raw tables and the people or tools querying them.
- Server-Sent Events (SSE)
A web standard where the server keeps one HTTP response open and pushes a stream of text events to the client as they happen.
- Serverless
A model where the cloud provider runs and scales the servers for you, and you pay only for what you use, with no machines to manage.
- Service account
A non-human Google Cloud identity used by an application, job or machine, with its own permissions, rather than a person's account.
- service control policy
An AWS Organizations policy that caps the permissions available in accounts below it, whatever their own IAM grants say.
- service worker
A script the browser runs in the background for a website, which can cache its files so the site works offline.
- session cookie
A small value the browser stores and sends back with each request, so a site remembers that you have already signed in.
- Session state
The small key and value store an agent framework such as ADK keeps for one conversation, so tools, callbacks and turns can share information.
- SFTP
SSH File Transfer Protocol, a common way to copy files securely to and from a server.
- Shadow mode
Running a new version on copies of real traffic alongside the current one, logging its answers for comparison while users still get the current version's answers.
- Shuffle
The step where a distributed engine such as Spark moves rows between machines so rows sharing a key end up together, for a join or grouping.
- side effects
Changes a step makes outside itself, such as starting a job or writing to a database, beyond just returning a result.
- signed URLs
Web addresses carrying a time-limited signature that grants temporary access to one Cloud Storage object without any other credentials.
- SKUs
Stock keeping units: the unique codes a retailer gives each distinct product variant, such as one boot model in one size.
- SLA
Service level agreement: a provider's written commitment to a level of service, such as uptime, often with credits if it is missed.
- softmax
A function that turns a list of scores into positive weights that add up to 1, giving the largest share to the highest score.
- source ports
Numbered slots on an IP address that keep outbound connections apart; each connection a VM opens through Cloud NAT uses one.
- Span
One timed step inside a trace, such as a single model call or tool call, with its start time, end time and details.
- split loads
Shipping one order in several smaller deliveries instead of one full truck, which costs more per unit.
- SQL
Structured Query Language: the standard language for asking a database for data, for example to select spend for one brand.
- SQL injection
A security hole where text typed into an app is run as part of a database query, letting an attacker read or change data they should not reach.
- SQL warehouse
A Databricks compute resource built for running SQL queries, dashboards and BI tools.
- SQLite
A small database engine that keeps a whole database in one ordinary file, with no separate server to run.
- SRAM
Very fast memory built into the chip itself, far smaller than HBM, used to keep the data the chip needs right now close to the maths.
- SRE
Site reliability engineering: Google's discipline of running production systems with software engineering methods, described in its public SRE books.
- StableHLO
An open, portable format for describing machine learning computations, used to hand a model from a framework such as PyTorch or JAX to a compiler such as XLA.
- staged rollout
Releasing a change to a small group of users first and widening it step by step, so problems are caught before everyone is affected.
- staging
A copy of the system set up like production but not used by real users, where changes are tried before release.
- Standard PayGo
Standard pay as you go: paying per token with no reserved capacity, so your requests share Google's capacity with other customers.
- stateless
Describes a service that keeps no memory of earlier requests itself, storing any data elsewhere, so any copy of it can handle any request.
- static site
A website made of pre-built files served as they are, with no server-side program creating pages for each visitor.
- stdout and stderr
Standard output and standard error: the two text streams every program writes to by default, for normal output and for error messages.
- stockouts
Times when a product has run out at a store or warehouse, so customers cannot buy it.
- Storage credential
A Unity Catalog object that wraps a cloud identity, such as a Google service account or an AWS IAM role, which Databricks uses to reach a storage bucket.
- Streamable HTTP
The current MCP transport, in which the client sends requests over ordinary HTTP and the server can stream its replies back on the same connection.
- sub-processors
Other companies a processor hires to help handle the customer's personal data, such as a data centre or support provider.
- Subnet
A range of IP addresses carved out of a network, usually written in CIDR notation; for example, a /28 range holds 16 addresses.
- supervised fine-tuning
Further training a model on example inputs paired with the desired outputs, so it learns a particular task, format or style.
- sycophancy
A model's tendency to tell users what they seem to want to hear, such as agreeing with a wrong claim, instead of what is true.
- synchronous
Done while the caller waits for the result, as opposed to submitted now and collected later.
- synthetic data
Training examples generated by a model or program instead of collected from real people, used when real examples are rare, sensitive or costly.
- SynthID
Google's technology for adding an invisible watermark to AI-generated content, such as images, that software can later detect.
- system prompt
The standing instructions sent to the model on every call before the conversation; in ADK this is the agent's instruction.
T
- tau-bench
A 2024 benchmark from Sierra that tests agents on realistic multi-turn customer service tasks with tools, simulated users and policy rules.
- TCP
Transmission Control Protocol: the internet protocol that delivers data reliably and in order, resending anything lost along the way.
- telemetry
Data a running system sends out about itself, such as traces, logs and metrics, so people can see what it did.
- temperature
A setting that sharpens or flattens a model's probabilities over possible next tokens; a higher value spreads probability more evenly across options.
- Terraform
An open source infrastructure as code tool: you describe cloud resources in configuration files and it creates or changes them to match.
- thinking tokens
Tokens a reasoning model spends working through a problem before it answers; they are billed as output although the user never sees them.
- three-legged OAuth
An OAuth flow with three parties, the user, the app and the provider, in which the user signs in and approves what the app may do.
- time to first token
How long a user waits from sending a question until the first word of the answer appears.
- time travel
A Delta Lake feature that lets you query or restore an earlier version of a table, using its transaction log.
- TLS (Transport Layer Security)
Transport Layer Security: the encryption behind HTTPS, which also proves a server's identity with a certificate.
- TLS certificate
A digital certificate that proves a server's identity and lets clients set up an encrypted HTTPS connection to it.
- Token (credential)
A short-lived digital credential a program presents with each request to prove who it is and what it may access, like a temporary pass.
- Tokens (text)
The small chunks of text, often a word or part of a word, that a model reads and writes; context sizes, usage and prices are counted in them.
- topologies
The shapes in which agents are wired together, such as one after another, in a loop, or side by side.
- TPUs
Tensor Processing Units: Google's custom chips designed to train and run machine learning models quickly.
- Trace
A record of every step one request went through, such as each model call and tool call, with timings, used to see where time and errors went.
- tracing
Recording each step of a request, such as model calls and tool calls, with timings, so you can see what happened and where it slowed or failed.
- trajectories
The ordered sequence of steps an agent took to reach an answer, such as which tools it called, with which arguments.
- Transformer
The neural network design behind modern language models, built from repeated layers that each mix information across the words in the input.
- transit hub
A central network that other networks connect to, so traffic between them passes through one managed place.
- Trino
An open source distributed SQL engine that queries data where it sits, across many sources, without loading it first.
- trust store
The list of certificate authorities a device or program trusts; a certificate signed by anyone else makes the connection fail or shows a warning.
- try/catch
A programming construct that runs some code and, if it throws an error, jumps to a recovery block instead of crashing.
- tuple
A fixed, ordered group of values in Python, written in round brackets, such as (a, b).
U
- UDP
User Datagram Protocol: a lightweight internet protocol that sends data without checking delivery, used where speed matters more than completeness, such as video calls and DNS lookups.
- Unit test
A small automated test that runs one piece of code, such as a single function, on its own with known inputs and checks the output.
- Unity Catalog
Databricks' governance layer for data and AI: it registers catalogs, schemas, tables, files and models in one place, enforces who may read or change them, and records lineage and access.
- Unix time
A count of seconds since midnight UTC on 1 January 1970, the usual way computers store timestamps.
- UTCP
Universal Tool Calling Protocol: an open specification for describing how to call an existing tool or API directly, without a wrapper server in between.
- UUID
Universally unique identifier: a long random ID, such as one stored in a browser, that can single out one device across visits.
V
- vanishing gradient problem
When training a deep or recurrent network, the corrections reaching early layers shrink towards zero, so those layers barely learn.
- variance
How much a set of numbers spreads out; explaining 65% of it means the rebuilt numbers capture 65% of how the real numbers vary from token to token.
- vCPU (virtual CPU)
Virtual CPU: one share of a physical processor, on Compute Engine usually one hardware thread, that clouds use to size, bill and set quota for machines.
- vector
An ordered list of numbers, such as an embedding, that can be treated as a point in space so distances can be compared.
- vector databases
Databases that store embeddings, lists of numbers representing meaning, and find the items closest in meaning to a query.
- version control
A system such as Git that records every change to files, who made it and when, so any earlier version can be restored.
- versioned
Given a version label, such as v1 or v2, so old and new definitions can be told apart and changed deliberately.
- Vertex AI
Google Cloud's machine learning and generative AI platform for building, tuning and serving models such as Gemini; its services now form part of Gemini Enterprise Agent Platform.
- view
A saved query that looks like a table: reading it runs the query over the underlying tables, without storing a separate copy.
- virtual environment
An isolated folder of Python packages for one project, so its library versions do not clash with other projects on the same machine.
- vLLM
A popular open-source server for running large language models efficiently on GPUs, TPUs and other accelerators, with an OpenAI-compatible API.
- VM (virtual machine)
Virtual machine: a software computer, with its own operating system, CPU, memory and disk, that runs on a share of a physical server.
- VPC network
Virtual Private Cloud: a private, isolated network inside Google Cloud; one VPC network is global, spanning every region, with each subnet in a single region.
- VPC Service Controls
A Google Cloud feature that draws a security perimeter around chosen projects and services, so data in services such as Cloud Storage cannot be copied outside it.
- VPN
Virtual private network: an encrypted tunnel over the public internet that links a laptop or an office to a private network as if it were directly connected.
- VS Code
Visual Studio Code: Microsoft's free, widely used code editor, which hosts AI coding assistants such as GitHub Copilot.
W
- watermark
In stream processing, a moving time threshold that sets how late an event may arrive and still be handled, so older state can be dropped.
- watermarking
Hiding a signal inside generated content, such as a pattern in an image's pixels, that software can later detect to show the content was AI-generated.
- webhook
A web address that one system calls automatically to tell another that something has happened, such as a new order, so it can act on it.
- WebSockets
A web protocol that keeps a two-way connection open, so the browser and server can both send messages at any time.
- weights
The billions of numbers inside a model that are adjusted during training; together they hold everything the model learned.
- worker nodes
The machines in a Kubernetes cluster that actually run your containers, under instructions from the control plane.
- Workforce Identity Pool
A Google Cloud group of staff identities that sign in through an outside identity provider, such as Microsoft Entra ID or Okta, instead of Google accounts.
X
- XML
Based on XML, a text format that marks up data with nested, named tags, widely used in documents and older web standards.
Y
- YAML
A plain text format for configuration files that uses indentation to show structure, readable by people and programs.
Z
- Zero Data Retention
An OpenAI arrangement, for approved customers, under which API inputs and outputs are not stored after the request is processed.
- zero-shot
Using a pretrained model on a new task or dataset with no extra training on it.
- zones
Separate data centre locations within one cloud region, each with independent power and networking, so a failure in one does not take down the others.