Type-checked Worker API examples

The learning catalog is in examples. See the setup guide, Runtime examples and database integrations.

graph-loop-worker.ts, runtime/graph-loop.ts and runtime/placement.ts execute at module load. The per-API function files below only export functions; importing them does not issue requests. Supply configured application model/database/MCP adapters, then call the function you need. Model calls can consume provider credits, and write/update/delete examples perform those operations.

File	Reference	Resources
context.ts	CONTEXT	Explicit Context, Redis or ContextStateStore
memory.ts	MEMORY	MemoryResources; filters/cursors/orderBy belong to the adapter
infer.ts	INFER · Providers	InferClient/ModelConfig or configuration loaded by setupInfer
interaction.ts	INTERACTION	Workspace access; connected MCP client and allowed absolute paths
retrieval.ts	RETRIEVAL · Providers	Search/embedding/rerank backends; complete MemoryItem candidates or mapOutput

Run npm run typecheck at the repository root. tsconfig includes these files, while emitted package builds exclude them. These are usage examples, not alternative SDKs or automatic live integration tests. Keep each // example: region and the matching API reference aligned when changing signatures.

For an executable local flow, use npm run example:agent or npm run example:tools. MCP commands are in the setup guide. Load .env explicitly with Node --env-file=.env or your application loader. loadRuntimeConfigFile reads YAML and uses the supplied environment; it does not read .env.

Function guide

Objects and constructors below do not automatically execute Graphs. Model/database/MCP parameters are configured application-owned resources.

memory.ts

Function / object	Behavior
setupMemory	Load configuration and create a MEMORY Worker/Runtime and direct SDK using application-owned store/search resources.
getMemory	Read exact IDs/keys, deduplicate repeated IDs and check NodeResult success.
queryMemory	Read up to two pages ordered by id, pass nextCursor and merge items; the adapter must support this order field.
searchMemory	Use the storage plugin's default search strategy; return full memories and optional scores.
writeMemory	Write a language preference and return its database ID; repeated-key behavior belongs to the adapter.
updateMemory	Replace content and clear metadata by an existing id, demonstrating partial-update semantics.
deleteMemory	Delete IDs and inspect actual deletions; repeated IDs are not reported twice.
executeMemory	Call MEMORY.QUERY through the common execute entry point.
adaptDatabase	Wrap a compatible application adapter as store/search while preserving method this binding.
memoryErrors	Show structured failures for invalid limits and application MemoryError.
memoryGraph	Run MEMORY.GET through a Graph and close Runtime.
customGet	Define a node descriptor returning NOT_CONFIGURED without accessing a database.

infer.ts

Function / object	Behavior
setupInfer	Load YAML/environment configuration and create a Worker, Runtime and direct SDK with a shared cache.
sample	Sample once and inspect Message, finishReason and usage.
sampleActions	Declare read_text to the model; generate a request without executing a file read.
trajectory	Run a CoT trajectory and check both outer success and inner completed status.
strategyRequests	Construct CoT, Long CoT, ToT, GoT and self-consistency requests without model calls.
reflectModes	Evaluate or revise a candidate using critique/verify/revise mode.
deliberateModes	Process candidates using select/merge/consensus/debate; selectCount applies only to select.
cacheApis	Write and read cache entries, then invalidate by key/tag/namespace.
executeInfer	Sample through execute with a per-call timeoutMs.
streamApis	Consume all four reasoning streams and terminal outputs; this makes multiple model calls.
cancelInfer	Pass an already-aborted signal and observe cancelled without issuing a model request.
cacheProviderApi	Call cache write/lookup/invalidate directly; return raw outputs without NodeResult.
refineStrategy	Define a strategy that drafts, revises, deliberates and records public decision steps; inject before use.
providerRegistryApis	Register, resolve, invoke and unregister a model provider.
providerStream	Consume provider streaming, falling back to invoke when streaming is unsupported.
httpModelProvider	Construct an HTTP provider from supplied configuration; construction does not send requests.
sampleDescriptor	Wrap SDK sample as a NodeDefinition without automatic registration/execution.

interaction.ts

Function / object	Behavior
readTextTool	Define a workspace reader with argument validation, capability metadata and Sandbox access.
consoleSink	Define a console sink that prints messages and returns an accepted receipt.
setupInteraction	Register a file tool, connected MCP client and output sink with a Worker/Runtime.
interactionNodes	Compose registries, createInteractionNodes and defineWorker.
interactionHandlers	Compose the four Interaction nodes using handler factories.
toolRegistryApis	Register, list allowed tools, invoke and unregister inside a real WorkerContext.
invokeTool	Read README.md through Runtime; ExternalResult has no output wrapper.
mcpRegistryApis	Discover and invoke read_text_file with McpRegistry, then unregister the client.
invokeMcp	Discover and read through Runtime while distinguishing both response shapes.
observeApis	Compare pure normalization with OBSERVE using supplied outcomes, without external requests.
outputApi	Send structured messages and artifact references to the console sink; this does not create artifact files.
missingRecordTool	Define a NOT_FOUND tool outcome to illustrate business failure.
rejectedSink	Define a sink returning a rejected delivery receipt.
interactionGraph	Read README.md and package.json through Graph + Loop and output two observations.
observationDefinition	Define an OBSERVE node descriptor without registration/execution.
mcpClientAdapter	Wrap a neutral MCP client, preserving this binding and optional fields.

retrieval.ts

Function / object	Behavior
request	Define request data: agent memory query, kb target, limit=5; no execution.
setupRetrieval	Register the optional Worker and logical target/strategy and create the direct SDK.
retrievalSearch	Search the same backend through SDK and Runtime and inspect candidates.
registryResolve	Resolve a target's default strategy provider and inspect raw results.
cancelRetrieval	Pass AbortSignal as search's third argument and fallback defaults as the second.
embeddingApis	Compare embed with batched embedContents and validate vector dimensions.
httpEmbedding	Create a real HTTP embedding provider from environment and use YAML batch configuration.
externalVector	Embed externally before passing the vector to database search.
nativeVector	Send the raw query to a backend with built-in embedding.
precomputedVector	Search using an existing vector after dimension validation.
textSearch	Wrap native database text search without embedding.
hybridSearch	Run vector and keyword branches concurrently and combine with weighted RRF.
rerankApis	Create a cosine reranker from embeddings and compare indexes/scores with mapped candidates.
rerankSearch	Expand the recall pool, rerank with an injected provider and return the requested count.
sqlSearch	Create a PostgreSQL full-text adapter with injected query and bound text/tenant parameters.
milvusSearch	Adapt Milvus requests/responses with injected SDK search, namespaces and complete MemoryItem records.
nativeMemoryInRetrieval	Convert a native MemorySearchProvider into a retrieval provider and back into Memory results.
nativeMemoryWithNamespace	Map Retrieval namespace to the adapter's tenant filter; this does not replace authentication.
localMemoryPipeline	Reuse a retrieval pipeline inside MEMORY without starting a RETRIEVAL Worker.
delegatedMemory	Delegate MEMORY.SEARCH to Runtime RETRIEVAL.SEARCH using same-process routing.
mapTextCandidates	Map complete candidate business data to MemoryItem; ID/snippet indexes need batch hydration.
retrievalDescriptor	Wrap the retrieval SDK as a NodeDefinition while retaining SDK validation and NodeResult.

context.ts

Function / object	Behavior
setupContext	Create the direct SDK and Worker; see CONTEXT for configuration.
loadContext	Load messages, items and references.
selectContext	Select context for inference or memory.
updateContext	Apply incremental changes and preserve provenance.
compressContext	Compress against a budget.
executeContext	Use the common execute entry point.
cachedContext	Invoke all four Context nodes in cache mode.
redisStore	Configure Redis storage and explicit versions.
contextServices	Inject custom Context services.
ragContext	Configure a RAG selection strategy.
contextRetrieval	Reuse a RETRIEVAL provider for Context.
contextToInfer	Convert selected Context for INFER.
memoryToContext	Compose Memory and Context in a Graph.
contextFlows	Run predefined RAG and Skill flows.
contextErrors	Demonstrate Context failure paths.
directStrategies	Reuse built-in strategies directly.
customContextLoad	Define a custom load node descriptor.
contextToMemory	Write selected content into long-term MEMORY.
toolToCachedContext	Write tool observations into cached Context.
remoteContextRetrieval	Delegate search to a separate RETRIEVAL Worker.

Suggested order

1. Start with the complete examples without optional services.
2. Choose a function for a Worker and inject resources matching its signature. Close Runtime returned by setupMemory/setupInfer/setupInteraction/setupRetrieval.
3. Run npm run typecheck after changes. Running node docs/worker-api/examples/memory.ts alone does not call its exported functions.

Database integrations cover Redis Context and SQL/Milvus Memory. Runtime examples cover a first Graph, custom Workers, events, artifacts, cleanup and RAG/Skill/Tool/MCP/ReAct flows. The per-function modules are safe to import without issuing requests; consult each directory for executable entry points.