Knowledge

A knowledge base is a corpus your operators can search and cite. Drop in docs, FAQs, policies, meeting notes, customer transcripts — operators query via the queryKnowledge tool and get back ranked snippets with citations.

Two engines

The knowledge layer is pluggable. Two engines are available today; the operator's query is the same shape regardless of which engine is wired.

GBrain — hybrid graph + vector

Self-wiring knowledge graph + hybrid vector/keyword search. Runs inside the guild's workstation sandbox; persists across pause/resume via the GCS backup. Best for:

• Chronological notes (what happened when)
• People and companies (relational queries)
• Decisions (what we chose, what the alternatives were)

Requires provider keys (ANTHROPIC_API_KEY, OPENAI_API_KEY) — supply your own or use guild credits.

Qdrant — vector-only

Bring your own Qdrant. Text + URL ingest, OpenAI-compatible embeddings (text-embedding-3-small by default), vector search. Pair with a managed Qdrant cluster, qdrant.tech, or a self-hosted Docker container.

Best for: pure semantic search across long-form documents, citation-heavy RAG, simple corpora.

Activating an engine

On the Available Engines card:

1. Pick the engine (GBrain or Qdrant).
2. Fill the required fields (provider keys, Qdrant URL, etc.).
3. Click Activate.

The activated engine becomes available under Active Knowledge Bases; operators with the queryKnowledge action can hit it.

How operators query

Operators see the same tool surface regardless of engine:

queryKnowledge({
  base: "support-faqs",        // optional — defaults to the guild's default
  query: "refund policy for annual plans",
  k: 5,
})
// → [{ text, score, citation: { source, span } }, …]

Citations are surfaced inline in conversation replies — the user clicks a citation and lands on the source document in the workspace.

Ingestion

Two paths:

1. ingestKnowledge tool — operators ingest content into the knowledge base as part of their workflow ("save this customer transcript", "index this competitor analysis").
2. Workspace file → knowledge — drop a file in a workspace and use ingestFromWorkspace(path) to index it.

GBrain auto-extracts entities and relationships as it ingests; Qdrant just chunks + embeds.

Switching engines

The engine binding lives on the knowledge base, not in the operator's tool calls. You can swap GBrain → Qdrant without changing any operator prompt — the tool signature is uniform.

What's next

• Actions → Knowledge — the queryKnowledge / ingestKnowledge / listKnowledgeBases actions
• Workspaces — where the source files live
• Operators — who calls the knowledge tool