Platform · Knowledge base
Agent answers grounded in your content, not the model's guess.
Load documents, pages, and policies into a knowledge base once. Agents retrieve from it at answer time across voice, SMS, and web chat, and are instructed to say when the answer is not there rather than guess.
Get started for free. Pay as you go. Or talk to sales
What it does
Three source types, one vector index
Add pasted text, uploaded files, or URLs to a knowledge base. Content is split into chunks and embedded for vector search, and the dashboard tracks each source from pending through complete, with chunk counts and error messages when processing fails.
- Source types: pasted text, file, URL
- Per-source status: pending, processing, complete, error
- Chunk counts and error messages shown per source
Retrieval tuned per agent
Attach one or more knowledge bases to each agent. Each attachment sets its own top-K, from 1 to 20 matches with a default of 5. Chunk size, overlap, and embedding model are set per knowledge base at creation, so retrieval granularity matches your content.
- Top-K adjustable from 1 to 20 per attachment, default 5
- Chunk size 100 to 4,000 tokens, overlap up to 1,000
- Two embedding models, text-embedding-3-small default
One index behind chat, SMS, and voice
Web chat and SMS turns run retrieval before the model answers, and matched passages are injected into the prompt for that question only. Voice agents query the same knowledge base mid-call through a built-in retrieval tool, so phone answers draw on the same content as text.
- Retrieval runs on every direct chat and SMS turn
- Voice agents call the knowledge base as a mid-call tool
- Chat retrieval queries attached bases in parallel
Prompt-injection hardening built in
Retrieved passages enter the prompt wrapped in tags and marked as untrusted reference data. The agent is instructed to use them only as source facts, treat embedded instructions as data rather than commands, and say it does not have the answer instead of inventing one.
- Passages are marked untrusted reference data
- Embedded instructions are treated as data, not commands
- Agents may not invent facts absent from the passages
Follow-up questions keep their subject
On web chat, the retrieval query stitches the last four turns of the conversation, so a bare follow-up like 'and the price?' still searches against the right subject. Assistant turns are capped in the query so the customer's actual question drives the match.
- Web chat retrieval includes the last 4 conversation turns
- Assistant turns capped at 300 characters in the query
- Full retrieval query capped at 1,000 characters
Failure never blocks a reply
Retrieval is best-effort by design. If the vector store or embedding call fails, the turn proceeds without grounding instead of erroring. On voice, a failed lookup returns a readable 'knowledge base unavailable' message the agent can work with rather than a dropped tool call.
- A failed retrieval never blocks the agent's reply
- Voice tool failures return a message, not a broken call
- 8-second timeout on query embedding
At scale
Built for the agent lifecycle.
01 · Deploy
Create a knowledge base, add text, file, or URL sources, and attach it to an agent. The same content backs chat, SMS, and voice; the voice retrieval tool is provisioned automatically.
02 · Test
Query the retrieval endpoint with a test question and inspect matched passages and scores. Sandbox and production knowledge bases stay separate, so staged content never appears in live answers.
03 · Scale
Attach multiple knowledge bases to one agent; chat retrieval queries them in parallel and merges results by score. Top-K, chunk size, and overlap stay tunable as your content grows.
04 · Observe
Per-source status, chunk counts, and error messages show what is indexed and what failed. When retrieval finds nothing, agents say so instead of guessing, which makes content gaps visible in transcripts.
Under the hood
- SOURCES
Pasted text, files, and URLs, tracked per source through processing
- RETRIEVAL
Vector search; each knowledge base isolated in its own index namespace
- TOP-K
1 to 20 matches per attachment, default 5
- CHUNKING
Chunk size 100 to 4,000 tokens, overlap 0 to 1,000, configured per knowledge base
- EMBEDDINGS
text-embedding-3-small default; text-embedding-3-large selectable per knowledge base
- CHANNELS
One knowledge base answers web chat, SMS, and voice calls
- ISOLATION
Scoped to workspace and environment; production and sandbox stay separate
FAQ
How does retrieval actually work?
Each question is embedded and run as a vector search against every knowledge base attached to the agent, in parallel. Results are merged by score, deduplicated, and trimmed to top-K. The winning passages are injected into the prompt for that turn only, wrapped and marked as untrusted reference data.
What content can I load into a knowledge base?
Three source types: pasted text, uploaded files, and URLs. Each source is tracked through pending, processing, complete, or error, with a chunk count once indexed. Chunk size (100 to 4,000 tokens) and overlap are configurable per knowledge base.
What happens when the answer is not in the knowledge base?
The agent is instructed to say it does not have the answer and offer to help further, rather than invent one. An empty retrieval result is a valid outcome the agent reasons over, not an error.
Do voice calls use the same knowledge base as chat?
Yes. Chat and SMS retrieve inline before the model answers. Voice agents call the same knowledge base as a mid-call tool with a 10-second timeout, hitting the same index and content. If the lookup fails mid-call, the agent receives a readable message instead of a broken call.
How many knowledge bases can one agent use, and what does it cost to try?
An agent can attach multiple knowledge bases, each with its own top-K from 1 to 20. All attached bases are queried in parallel on every direct chat or SMS turn. Signup is free pay-as-you-go, so you can load content and verify retrieval before running live traffic.
