Platform · Knowledge base
It answers from your documents — or says it can't.
Upload the price list or point at a page on your site — then ask. When your documents hold the answer, callers hear it, with its source; when they don't, the agent doesn't improvise — it takes the fallback you drew in the flow.
“Is a cleaning covered by my insurance?”
Services guide
And when the answer isn't in your documents, it says so — it doesn't invent one.
From upload to answer
- PDF · DOCX · CSV · TXT · URL
Five ways in
Files are extracted by type; URLs are fetched through an outbound guard that blocks unsafe destinations before a byte is read. Every document carries its status honestly — processing, chunked, embedded, or error with the reason.
- CITED-BY-NAME
Every answer names its source
Every answer can say which document it came from — by name. Behind that, a recursive chunker splits your text into overlapping retrieval-sized pieces, each one carrying its source document's name; the sizes are on the spec sheet below.
- ALL-OR-NOTHING
No half-indexed documents
Chunks are inserted in a single statement: either the whole document becomes retrievable or its status reads error — never a Ready badge over zero retrievable chunks. A retried ingestion never duplicates chunks.
- SIMILARITY ≥ 0.70
Evidence, or a decline
If the evidence in your documents isn't strong enough, it doesn't guess — it takes the fallback you drew: a message or a transfer, never an improvised answer. Curated FAQ entries answer first; everything else must clear the similarity bar the spec sheet states.
- TENANT-GATED
Your knowledge stays yours
Every query is gated on knowledge-base ownership. A foreign organization's probe gets an empty result set — not an error that leaks the base even exists.
Spec sheet — as enforced in code
- PDF · DOCX · CSV · TXT · URL
Ingestion formats
- 500 / 100 tokens
Chunk size / overlap
- similarity ≥ 0.70
Retrieval threshold
- similarity ≥ 0.85
FAQ match threshold
- all-or-nothing insert
Ingest contract
Thresholds are the defaults the retrieval functions enforce in the database — not tuning advice.
The platform