Features
Knowledge base
Drop your FAQs, product specs, or pricing tables into Floo and your agent will reference them on every call. No retraining, no fine-tuning - just upload and go.
How it works (RAG, plain English)
When a caller asks a question, the agent doesn't answer from memory. It searches your knowledge base in real time, pulls the 2-3 most relevant snippets, and uses them to construct an accurate reply. We call this retrieval-augmented generation (RAG) - the LLM does the talking, but the facts come from your documents.
Why this matters: the agent will tell a caller the right price, the right return-policy window, the right office address - and won't hallucinate. If the answer isn't in your KB, it'll say so and offer to escalate.
Three ways to add knowledge
Upload a file
PDF, DOCX, TXT, MD up to 25MB. We extract text, chunk it, and embed it.
Best for product manuals, policy docs, brochures.
Paste text
Just paste raw text into the editor. Use Markdown for headings if you want hierarchy.
Best for FAQs, scripts, hand-curated answers.
Add a URL
We crawl the page (single URL, no recursion) and re-fetch nightly.
Best for live pricing pages, status pages, public docs.
All three methods land in the same place - the agent doesn't know (or care) which method you used. You can mix them in a single KB.
Best practices
Do this
- • Upload FAQs, scripts, pricing tables - Q&A format works beautifully.
- • Keep individual sources under 50 pages. Smaller = better retrieval.
- • Use headings (
## Section) so chunks break on natural boundaries. - • Re-upload when content changes - embeddings don't auto-refresh on file uploads.
- • Test by asking the agent a known-answer question in the browser preview.
Avoid this
- • Dumping a 500-page PDF - search quality degrades, costs climb.
- • Uploading PII, customer records, or credentials. KBs aren't access-controlled.
- • Image-heavy PDFs without OCR - we extract text only, no vision yet.
- • Long-form prose with no headings. Chunking falls back to character splits.
- • Scraping URLs you don't own - respect robots.txt and ToS.
Chunk size - what to know
Floo splits each source into chunks of ~500 tokens (about 300-400 words) with 50-token overlap. You don't configure this. What you can control is how the source is structured: documents with clear headings produce semantically clean chunks; a wall of text produces noisier ones. If retrieval quality feels off, add headings.
How the agent uses the KB during a call
On every caller turn, the agent embeds the question, runs a cosine-similarity search against your KB, and stuffs the top-3 chunks into its context window before generating a reply. Total added latency: ~120ms - barely noticeable.
You can see exactly which chunks the agent retrieved on the call detail page under the Knowledge retrievals tab. Great for debugging “why did the agent say X?” moments.
Ready to add your first KB?
Open any agent → Knowledge tab → Add source. Five minutes and your agent stops hallucinating.
Open agents