Assistant Knowledge (/app/assistants/{id}/knowledge) ingests files and websites into pgvector chunks. Embeddings use local sentence-transformers/all-MiniLM-L6-v2 (not OpenAI embeddings). Knowledge nodes on the flow canvas and runtime fallback path query these chunks.
Sources
- Upload PDF, DOCX, TXT, or website URLs — max file size from knowledge_max_file_size_mb.
- Per-source status — indexing, ready, failed with error_message.
- chunk_count, embedding_model stored on KnowledgeSource.
- Intent bindings — attach sources to specific intents for scoped retrieval.
- Test retrieval — query box with combined vector + keyword scores.
Ingestion pipeline
- Parse document or crawl URL.
- Chunk with overlap (knowledge_chunk_size / knowledge_chunk_overlap).
- Embed with MiniLM; store in knowledge_chunks with tsvector keywords.
- Synthesis optional — synthesize_answer uses extraction LLM when building answers from chunks.
AI Agent Knowledge (/app/agents/{slug}/knowledge) is a separate RAG store for the agentic runtime — same embedding approach, different API namespace.