Skip to main contentThe Knowledge Base allows you to upload documents that agents can query during a conversation. This is essential for providing accurate, business-specific information that isn’t contained in the LLM’s base training data.
How it Works
- Ingestion: You upload a file (PDF, TXT, DOCX) or raw text.
- Indexing: We split the document into chunks, generate vector embeddings, and store them in a vector database (Pinecone).
- Retrieval: When a user asks a question, the agent performs a semantic search to find relevant chunks.
- Generation: The relevant chunks are injected into the LLM’s context window, allowing it to answer accurately.
Managing Documents
- PDF
- Text files (.txt)
- Microsoft Word (.docx)
- Raw text input
Attachment
Documents are not automatically used by all agents. You must explicitly attach a document ID to an Agent ID. This allows you to have different knowledge sets for different agents (e.g., “Sales Agent” vs. “Support Agent”).
Best Practices
- Chunk-friendly formatting: Use clear headings and short paragraphs.
- Specific information: RAG works best for facts (prices, policies, hours).
- Update frequency: If information changes, you must delete the old document and upload the new version.
