Skip to content

Glossary ​

Every kdeps term defined in one place. The first mention of any term on other docs pages links here. A workflow is the one way to run kdeps; the agent loop is a chat with loop: inside it.

TermDefinitionMore
actionIdA unique string that identifies a resource within a workflow. Used as the target of requires: dependencies.Resources overview
agentA workflow whose chat has a loop: block: the model calls tools over many rounds and keeps one conversation per loop.session id. There is no kind: Agent - kind: is Workflow, Component, or Agency.Agent loop
input sourceHow requests reach a workflow: api (HTTP), bot, file, or interactive (a terminal prompt where each line is one request). Set under deploy.input.sources.Input sources
agencyMultiple agents composed into one system. One agent delegates a task to another via the agent: action type; the callee runs its full pipeline and returns its output.AI agencies
apiResponseA resource action type that returns a structured JSON response to the client. Usually the terminal node of a workflow.API response
before / afterExpression blocks that run before or after a resource's main action. Used for data preparation, normalization, and validation. Statements execute in order.Expression blocks
chatA resource action type for LLM chat completion. Sends a prompt and returns the response. The most common action type.LLM (chat)
checkA list of boolean expressions in validations.check. All must be true or the resource fails and the workflow stops.Validation and control flow
componentA reusable, packaged resource bundle. Installed from the registry or built in a components/ directory. Declared as kind: Component.Components
tools.componentsTools provided by a component that are exposed to the calling agent or workflow for the LLM to invoke.Components
deterministicSame input, same output, every time. In kdeps this describes the workflow-mode pipeline - route matching, requires: ordering, validations, and response shaping all resolve the same way for a given request. It does not describe the LLM's text output. Contrast probabilistic.Deterministic by design
embeddingThe shipped component around the native embedding executor: a local SQLite keyword store (index, search, upsert, delete), an embedding backend (vectorize, embed_query), and Cohere rerank. Keyword matching is SQL LIKE, not vector similarity.Embedding
execThe kdeps component that runs a shell command through the kdeps API and returns the structured result (success, exitCode, stdout, stderr, command, result, timedOut).exec
evalThe judge panel. A chat turns it on with judges: or judges: auto under guardrails. It reviews the answer on its own.Judge panel
exprShort for expression: a statement evaluated by the expr-lang engine. Used in before: / after: blocks, validations, and string interpolation.Expressions
file()An expression function that reads uploaded or local files. Takes a glob pattern and an optional selector (first, last, count, all, mime:<type>).Unified API
fileA kdeps component (component: file) for filesystem operations: read, write, patch, list, delete, mkdir, copy, move, append.File
get()The primary data access function. Reads from query params, headers, body, session, memory, environment, or resource outputs, with auto-detection.Unified API
gitA kdeps component (component: git) for version control: status, diff, log, commit, branch, push, pull.Git
graderA check attached to a judge with graders:. Kinds: exact, contains, not_contains, regex, json, expr. It runs on the answer in process. A failure rejects that judge before it calls the model.Judge panel
http_requestA kdeps component (component: http_request) for HTTP requests. Supports all methods, headers, query params, body, retry, caching, and named connections.HTTP client
info()An expression function that returns request metadata: ID, timestamp, path, method, IP, sessionId, filecount, files, filetypes.Unified API
itemsAn array a resource iterates over. The resource runs once per item. Access the current item via item.current() or get('current').Items iteration
harness.output.jsonA boolean field on chat resources. When true, forces the LLM to return valid JSON with no markdown wrapping.LLM (chat)
judgeOne reviewer on the eval. The name stays judge: judges:, judge_verdict, the harness prompts. The items are filed on the evals shelf.Judge panel
kdeps APIThe process control listener of kdeps run. It binds 127.0.0.1, requires KDEPS_API_TOKEN, and serves discovery, tool calls (read_file, write_file, bash_exec, bash_job_list, bash_job_wait, sql_ops, memory_save, memory_search, memory_delete, memory_query, http_request, search_local, load_document, embedding, embedding_search, embedding_vectorize, the four approval_* tools, instruct, and the asset-declared tools (calculator, edit_file, file_ops, git_ops, google_cache_*, list_files, md5_file, sql_list_tables, sql_describe_table, sql_query, tail_file, web_search, web_scraper, wikipedia, the rerankers, exa_search, ocr_image, retrieve_context, transcribe_audio, transcribe_audio_offline, wolfram_alpha, zapier_*)). It is not the workflow apiServer and it is not /_kdeps/.kdeps API
loopA while-loop on a resource. The body runs while the while expression is true, up to maxIterations. every: turns it into a scheduled ticker.While-loop iteration
memory (expression)Request-scoped key/value storage. Set with set('key', value, 'memory'); read with get('key', 'memory'). Cleared when the request completes.Session and memory
memory (persistent)Project-scoped storage for the agent loop, kept as markdown notes in the knowledge vault, that survives across sessions. Built-in tools: memory_save, memory_search, memory_delete, memory_query. The graph is injected into the prompt. Workflow resources can still call memory_list().Persistent memory
memory graphA directed graph of memory entries linked by their References fields and auto-linked by type. Inlined into the <memory> block in causal order.Memory internals
output()An expression function that reads the output of a completed resource by its actionId.Unified API
probabilisticThe same prompt can produce different output on each call. Every language model - Claude, GPT, Gemini, Groq, Ollama, local llamafile / GGUF - is probabilistic. A kdeps chat: resource inherits this; the pipeline around it stays deterministic.LLM provider reference
pythonThe kdeps component that runs a Python script through the kdeps API and returns stdout parsed as JSON when possible. Supports inline scripts, file paths, packages, and virtualenvs.python
requiresA list of actionIds that must complete before this resource runs. Defines the DAG edges. Transitive dependencies resolve automatically.Execution flow
resourceThe fundamental building block of a workflow: an actionId, a primary action type, optional requires, and optional validations, loop, and error handling.Resources overview
scenarioA test case for a resource or workflow: inputs and expected outputs, checked during kdeps validate.Validation examples
search_localA kdeps component (component: search_local) that searches local files by keyword and glob, optionally with a TF-IDF index and graph boost.search_local
sessionHTTP session storage across requests from the same client. Set with set('key', value, 'session'). Not the agent loop conversation (loop.session).Session configuration
set()Stores a value in memory (this request) or session (across requests).Unified API
shelfThe engineering word an asset is filed under: model, context, harness, tools, memory, rag, agent, guardrails, evals, workflow, deploy. The path stays the id (harness/safety). kdeps list <shelf> shows one. Install profiles are a separate axis.Installation
skipA list of boolean expressions in validations.skip. If any is true, the resource is skipped silently and the workflow continues.Validation and control flow
sqlA kdeps component (component: sql) for SQL against PostgreSQL, MySQL, SQLite, SQL Server or Oracle. Parameterized queries prevent injection.SQL
streamingA boolean field on chat resources. When true, the LLM response streams token by token.LLM backends
targetActionIdThe entry-point resource of a workflow, set as targetActionId: under metadata:. Execution resolves the dependency graph backward from here.workflow.yaml
toolsFunctions registered with the LLM. On a chat resource, tools are the functions in tools:.Built-in tools. Function calling
validationsA resource-level block with skip, check, routes, methods, and error. Controls whether a resource runs and how it fails.Validation and control flow
workflowThe top-level unit of execution: a YAML file defining resources, dependencies, and configuration. Declared as kind: Workflow.workflow.yaml

See also ​

kdeps has the Apache 2.0 License.