Knowledge Base
The voice agent answers from your uploaded content. When a caller asks a question, the agent searches the knowledge base for matching chunks and quotes the exact wording. If nothing matches, the agent says “let me check that and have someone call you back” — and never invents an answer.
Supported file types
| Type | Notes |
|---|---|
| Text + scanned PDFs both work. Multi-page supported. | |
| TXT / MD | Plain text. Best for FAQ pages, policy docs. |
| CSV | First row is treated as headers. Each row becomes a “key: value” line per column. |
| DOCX | Word documents. |
| XLSX | Excel spreadsheets. |
| PNG / JPG | Vision model reads the image content (e.g. menu boards, price lists). |
Max file size: 50 MB.
Upload a file
On the agent edit page, go to the Knowledge Base tab → Upload. Drop a file or click to pick.
Status flow: processing → indexed (or failed with a reason).
Refresh the page or click the row to see the latest status.
Processing usually takes 5-30 seconds for typical files; large PDFs
or image-only files take longer.
Test a query
Use the Test query input to ask a question and see what chunks the agent would retrieve. Examples:
- “What are your opening hours?”
- “How much does a root canal cost?”
- “What’s the cancellation policy?”
You see the matching chunks + their similarity score. If a question returns no hits, the agent will say “let me check that and call back” on a real call — fix that gap by uploading a doc that covers it.
Best practices
- Split big docs into focused files. A single 200-page PDF works, but separate “menu.pdf” + “policies.pdf” + “team.pdf” gives sharper retrieval.
- Update + reupload when prices/hours change. The agent quotes what’s in the KB. Stale prices = unhappy callers.
- Use plain language. Marketing copy (“our award-winning experience”) is harder for the agent to ground from than direct facts (“we open at 9 AM Monday to Saturday”).
- Image content gets OCR’d — handwritten notes, faded scans, and busy backgrounds reduce extraction quality. Prefer clean digital PDFs when possible.
Removing files
Click Delete on any row. The agent stops referencing the file within seconds. Re-uploading isn’t supported in-place — delete the old, upload the new.
What the agent does NOT do
- Doesn’t read your CRM / sheets / Notion — only the files uploaded to this agent.
- Doesn’t browse the web — facts must be in the KB.
- Doesn’t share private content — every agent’s KB is scoped to that agent + that customer account.
Troubleshooting
| Symptom | Cause | Fix |
|---|---|---|
| Agent says “let me check that” on every question | KB is empty or failed | Check status on Knowledge Base tab. Re-upload failed files. |
| Wrong information surfaces | Stale doc still in KB | Delete the stale doc, upload the corrected version |
PDF status = failed: extraction failed | Image-only PDF too dark / busy | Convert to PNG and re-upload, or improve scan quality |
| ”Vision model failed” on image upload | Image is corrupt or too large | Resize to under 5 MB, save as fresh PNG/JPG |