Agent Knowledge (RAG)

Give an agentic worker its own knowledge corpus — a set of documents (PDFs, text) the worker retrieves from automatically. Before each LLM call the worker searches the corpus and injects the most relevant passages as a <knowledge> block in the system prompt. This is retrieval-augmented generation (RAG) as auto pre-retrieval: the model always sees grounded context, without you wiring a retrieval tool by hand.

This tier is not pluggable

The retrieval described on this page is Greentic’s built-in knowledge backend. It is selected at the runner build, not through a capability, so a third-party retrieval stack cannot replace it. To plug in your own RAG, see Add a custom RAG — it enters as a flow node or an agentic-worker tool instead.

How it works

1. Attach documents to the worker. They are extracted, chunked, and baked into the worker’s .gtpack as a corpus — the pack is self-contained.
2. At runtime, each turn: the worker embeds the query, searches the corpus for the top top_k passages, and prepends them to the prompt (after any long-term memory facts).
3. The model answers grounded in the retrieved passages.

AgentConfig.knowledge

"knowledge": {
  "provider":  "chronicle",                 // cap://dw.knowledge
  "embedding": { "provider": "openai", "model": "text-embedding-3-small" },  // cap://dw.embedding
  "top_k": 5
}

• provider — the knowledge store capability (cap://dw.knowledge), e.g. Chronicle.
• embedding — the embedding provider (cap://dw.embedding) used to index the corpus and embed queries.
• top_k — how many passages to retrieve per turn (default 5).

Runtime requirement

Knowledge retrieval is opt-in at the runner build, the same as long-term memory: the runner must be built with the knowledge-chronicle feature, and retrieval stays silently disabled unless the embedding environment is set. An unconfigured tier never fails startup — the worker just answers without retrieval.

Variable	Purpose
GREENTIC_KNOWLEDGE_EMBED_BASE_URL	Embedding API base URL.
GREENTIC_KNOWLEDGE_EMBED_API_KEY	Embedding API key.
GREENTIC_KNOWLEDGE_EMBED_MODEL	Embedding model.
GREENTIC_KNOWLEDGE_BACKEND	Graph/vector store (defaults to embedded SurrealDB).

Feature-gated runner build — gtc 1.1.2

The default deployed runner (features verify + agentic-worker) does not retrieve knowledge. gtc 1.1.2 ships a knowledge-enabled runner artifact (runner-full, built with knowledge-chronicle) and adds gtc start / gtc doctor checks that warn when a worker declares knowledge but the running binary or the GREENTIC_KNOWLEDGE_EMBED_* environment can’t serve it — instead of silently answering without retrieval.

deep_worker not yet covered

Automatic knowledge retrieval currently applies to the single_turn and agent_graph types. Wiring RAG into the deep_worker loop is planned but not yet available — a deep_worker will run, but it won’t auto-retrieve from an attached corpus.

Authoring

In the Designer, manage a reusable tenant knowledge library (upload and manage PDFs), then attach documents to a worker in the composer’s Knowledge section; the corpus is emitted into the worker’s pack.

CLI — gtc 1.1.2

From gtc 1.1.2, declare knowledge in a WorkerSpec:

knowledge:
  provider: chronicle
  embedding: { provider: openai, model: text-embedding-3-small, credential_ref: embed-openai }
  top_k: 5
  documents: [ ./kb/policies.pdf, ./kb/faq.md ]

gtc worker build reads the documents locally, extracts and chunks them, and bakes the corpus into the .gtpack.

See also

• Agent Types — single_turn, agent_graph, deep_worker
• Agent Memory — short-term and long-term memory