Playbook

It is allowed to say "I don't know"

Most document chatbots will answer anything, because a plausible answer looks like success. Playbook applies a grounding floor before it composes: the best-matching passage has to cover more than one distinct term from your question. A lone incidental word match — a page that happens to contain "policy" — doesn't qualify, and you get an honest miss instead, flagged as a possible gap in the knowledge base. Every answer that does come back carries the document and section it came from, so it can be checked.

Two products, one corpus

That second one is the difference between a search box and a colleague: the documentation doesn't just get read, it gets executed.

Keyless by default

Retrieval and answer composition run entirely at the edge with no model call and no per-token cost — the extractive composer stitches the most on-topic sentences from the top passages. Point it at an LLM and it writes more fluently, but it stays constrained to the retrieved excerpts; the anti-hallucination contract is the same either way. The index is rebuilt from storage per request, so a document ingested a second ago is searchable now.

Ingest, then ask

Markdown in: # is the title, ## starts a section, and a numbered list makes the document runnable.

# Refund Policy

## Eligibility
Customers may request a refund within 30 days of purchase. Digital goods are
refundable only if unopened.

## Processing a refund
1. Verify the order date is within the 30 day window.
2. Confirm the item condition with the customer.
3. Issue the refund to the original payment method.
4. Email the customer a confirmation.

Ask it anything — /ask is open:

curl -s 'https://playbook.forthecommongood.ai/ask?q=how+long+do+customers+have+to+request+a+refund'

Returns the answer, whether it was grounded, and the sections it came from with their retrieval scores.

Endpoints