knowledge

Base path: /api/v1/knowledge

The knowledge module is the brand-context and RAG layer that every agent reads from. You ingest URLs, files, and structured brand entries; 10ex chunks, embeds, and stores them in Qdrant; agents retrieve the relevant chunks at inference time. Resources include brand documents, ingestion jobs, structured brand entries (positioning, ICP, voice), and vector collections. Embeddings use OpenAI text-embedding-3-small. Schemas live in app/modules/knowledge/schemas.py.

What you can do with the knowledge API

• Ingest a URL (the worker scrapes, cleans, chunks, and embeds)
• Upload a file (PDF, DOCX, MD, TXT) for ingestion
• Maintain structured brand entries (mission, positioning, ICP, voice, do-not-say)
• Search the corpus directly with a semantic query
• Re-ingest a stale URL on demand
• Delete a document or a single chunk

Common operations

• Ingest a marketing site, then watch the job until embedding completes
• List documents grouped by source (URL, file, manual entry)
• Update a structured brand entry as positioning evolves
• Search the corpus to debug what an agent will retrieve
• Delete an outdated document so agents stop quoting it

Authentication & scoping

Workspace-scoped via API key in the Authorization: Bearer tnx_... header. Each workspace has its own Qdrant collection so retrieval never crosses tenants.

Pagination & filtering

Document and chunk list endpoints paginate cursor-style with filters by source type, ingestion status, and date. See Pagination.

How knowledge retrieval works at agent time

When an agent runs, it issues semantic queries against your collection, retrieves top-k chunks, and inlines them into the prompt with citations. If you don’t see the right info in an agent’s output, search the corpus directly to confirm the chunks exist and rank well.

MCP equivalents

Related

• Concepts: Knowledge and RAG for retrieval mechanics
• agents for the agents that consume this corpus
• Guides: Brand setup