Flocci Chat Start free
Use case

Flocci Chat for documentation and help centres

Short answer

Documentation search returns ten links; readers wanted one answer. Flocci Chat reads your docs — including versioned, MDX and JavaScript-rendered pages — and answers in prose with the source page cited, so a reader can verify in one click. New pages are learned the first time anyone opens them, with no re-index job to schedule.

The problem with docs search

Full-text search rewards the word the reader guessed, not the concept they meant. Someone typing "rate limit" finds nothing when the page says "throttling"; someone typing "how do I rotate a key" gets the changelog. Retrieval-augmented answering fixes the semantic half of that problem — hybrid vector plus full-text search finds the passage by meaning and by keyword, diversity reranking stops three chunks of the same section from filling the answer, and the reply quotes the page and links it. Readers still get the source; they just do not have to assemble the answer themselves.

Docs sites are exactly where crawler-only tools underperform

Modern documentation runs on Docusaurus, Nextra, Mintlify, VitePress and other client-rendered frameworks, and often behind version switchers and tabbed code panes. A server-side crawler frequently captures navigation chrome and misses the content. Flocci Chat's widget reads each docs page after rendering, so the version the reader is looking at is the version indexed — and when a page changes, the next visit replaces the old copy in the index automatically, keyed on a content hash.

The report you did not know you needed

Every question your docs failed to answer is captured as an unanswered intent. That list is the most honest documentation backlog in existence: real phrasing, real frequency, no survey bias. One click converts a recurring question into a canonical Q&A so the assistant answers it while you write the page properly. Teams that run this loop for a month usually find that five missing paragraphs were generating most of their support tickets.

Docs plus product, one workspace

Put the widget on the public docs with learning on, and inside the authenticated product with data-passive-ingest="off" so in-app screens are never indexed. Logged-in users can be verified with an HMAC signature, which lets the assistant answer plan-specific or account-specific questions through your own API instead of sending a developer to the billing page. Anything it should not answer — an escalating incident, a contract question — opens a ticket with the transcript attached.

Frequently asked questions

Can it answer questions about a specific docs version?

It indexes the pages your readers actually open, so versioned docs are captured per URL exactly as rendered. Ask version-specific questions the way your readers do, and cite the page you land on to confirm the right version was used.

Does it replace our docs search?

It complements it. Search is right when someone knows the page they want; an answering assistant is right when they know the problem but not the vocabulary. Both point at the same pages, and every answer links its sources.

What about code samples in answers?

Replies render full markdown — fenced code blocks, tables, lists and headings included — with links opened safely in a new tab. Chunking is section-aware and code-aware, so a sample is not sliced in half mid-function.

How do we keep internal runbooks out of it?

Only index what is public: the crawl covers your public docs, passive learning covers rendered public pages, and private surfaces carry data-passive-ingest="off". Internal runbooks simply never enter the knowledge base.

Try it on your own website

One script tag. 500 free credits — roughly 500 answered visitor messages. Every feature on, no card. The widget in the corner of this page is the product, running on its own site.

Create your workspace — free