Retrieval, not recall
Every answer is assembled from workspace-scoped API context and recent conversation — not the model's training memory.
AI support agent
Answers grounded in your API
An AI agent that answers only from your workspace-scoped API context, cites every source, and hands off when the evidence isn't there.
100% citedNo invented endpointsClarify or hand off
How do I authenticate the /orders endpoint?
Send a Bearer token in the Authorization header — keys are scoped per environment▍
docs/auth.md POST /v1/auth grounded
confidencehigh
The problem
A confident chatbot is worse than no chatbot
Generic AI support reads your docs and then improvises around the gaps — inventing endpoints, guessing auth rules, and stating made-up behavior as fact. For an API company, one hallucinated response can send a developer down a debugging path that never terminates. Trust erodes on the first wrong answer.
Why it works
Most support bots are confident and wrong.
Cited or it doesn't ship
Claims carry inline citations you can open. If the agent can't ground a statement, it says so and clarifies or hands off.
Injection-aware
Source content and user input are treated as untrusted. Redaction and careful mode keep prompt-injection attempts from leaking or misfiring.
answer · sources attachedcited
Rate limits are 100 requests/minute per key on the /v1 tier.
docs/limits.md GET /v1/orders every claim cited
confidencehigh
Cited in the open
You can see exactly why it answered
Every answer is assembled from your workspace context and carries the sources behind it. Open the citation, check the evidence. When the evidence isn't there, it clarifies or hands off instead of guessing.
Grounding
Answers built from evidence you can open
The agent retrieves the most relevant chunks of your API context, drafts an answer, and attaches the exact sources behind it. Reviewers and customers can click straight through to the endpoint, doc, or schema that backs each claim.
- Inline citations to endpoints, docs, and schemas
- Answers scoped strictly to your workspace context
- No claim without a source behind it
answer · sources attachedcited
Confidence
Low confidence clarifies — it never bluffs
The agent scores how well the evidence supports a question. Strong context yields a grounded answer; weak or missing context triggers a clarifying question or a clean handoff, with the reason recorded for the operator.
- Explicit confidence signal on every response
- Insufficient-context cases ask or escalate
- Careful mode raises the bar on sensitive threads
confidence · grade & route
Trust boundary
Safe with untrusted input and secrets
Ingested docs and customer messages are never trusted blindly. Credentials are redacted before they reach the model, prompt-injection patterns are contained, and operators get a sanitized debug trace that never exposes secrets or provider internals.
- Credentials redacted before every model call
- Prompt-injection resistant retrieval and prompting
- Operator-only debug traces, secrets stripped
trust boundary
How it works
Three steps, no heavy lift.
01step
Retrieve the evidence
The agent pulls the most relevant workspace-scoped context for the question at hand.
query → top-k context chunks
02step
Draft with citations
It composes an answer grounded in that evidence and attaches the sources it used.
answer + docs/auth.md + POST /v1/auth
03step
Grade and route
High confidence sends a cited reply; low confidence clarifies or hands off with a reason.
confidence: high → cited reply
100%
of answers cited
00
invented endpoints
auto
handoff when unsure
FAQ
Grounded AI agent questions
How does the agent avoid making things up?
It only answers from workspace-scoped API context and recent conversation, and it attaches citations to its claims. When the evidence isn't there, it clarifies or hands off rather than inventing endpoints, fields, or behavior.
Which AI provider or model does it use?
Provider and model routing is a platform concern handled on the backend. It isn't exposed as a customer-facing setting, so you get grounded answers without managing model plumbing.
Is it resistant to prompt injection?
Ingested content and user input are treated as untrusted. Credentials are redacted before model calls, injection patterns are contained, and the operator debug trace never exposes secrets or provider internals.
Can operators see why it answered the way it did?
Yes. Every reply carries a sanitized, operator-only trace showing the retrieved context and reasoning, with sensitive data stripped out.
Put a grounded agent on the front line
Let an agent that cites its sources — and refuses to guess — handle the first reply on every channel.