Knowledge base (RAG)

The knowledge base feature lets your AI agent answer questions using your actual docs, such as release notes, help articles, policies, anything text-based.

Supported formats

• PDF
• Markdown, plain text
• Word (.docx) · CSV
• JSON (extracted as text)

Upload flow

1. Go to Knowledge → Upload document
2. Pick a file (max 25 MB). It uploads directly to S3 via presigned POST
3. The worker picks up the job: extracts text → chunks at ~800 tokens → embeds with OpenAI text-embedding-3-small → stores in pgvector
4. Status flips to READY when done

Training Center

Manual single-file upload is only one way in. Open AI Agent → Training Center to feed the knowledge base several ways:

• Crawl a website: give it a URL and it fetches the pages, extracts the text, and ingests them just like an uploaded file
• Bulk file upload: drop in PDF, Word (.docx), and other text documents to parse and chunk automatically
• Q&A pairs: write them by hand, or import many at once from a CSV
• Playground: a sandbox chat to try the agent against your knowledge base before it goes live

How retrieval works

On every visitor message, the worker embeds the query and runs a cosine-distance search over your tenant's chunks. Top-K (default 5) chunks pass a similarity threshold and get inlined into the system prompt.

Platform-hosted AI, or your own embeddings

By default the platform hosts the AI for you, no key required: embeddings and replies run on managed keys, metered by real token usage, topped up automatically from credits, and handed to a human agent when the balance runs low. If you would rather self-host, inAI Agent → Embeddings you can set a custom embedding API endpoint, useful if you want to use Azure OpenAI, a local Ollama, or DuckLLM. The chat completion endpoint can be different from the embedding endpoint, because aggregator services often do not support embeddings.

Caching

Identical query text hits a 24-hour Redis cache, so the same FAQ doesn't re-embed on every visit. Anthropic prompt caching is applied to the system prompt + retrieved chunks for further cost savings.

When to skip RAG

Toggle RAG enabledoff if your AI agent is a pure small-talk / lead-qualification bot that doesn't need to cite docs.