Changelog

A living product.

Per-release notes, reverse chronological. The latest release is at the top. Subscribe via the RSS feed (coming soon) or follow @flowchat.

  1. Feature v0.1.0

    Launch — citation-grounded chat for your site

    The first public version of FlowChat. Crawl your site, build a knowledgebase your visitors can trust, and answer their questions with verifiable sources — or refuse when the answer isn’t on your docs.

    What you can do today:

    • Drop a chat widget on your help center in two lines of HTML. Works on any modern docs site — JavaScript-rendered, static, anything you can reach with a URL. Live in under an hour from sign-up.
    • Every answer cites a real page on your docs. Click any sentence in any reply and the page it came from opens. Your QA team can audit every conversation.
    • The bot refuses when it doesn’t know. Confidently-wrong answers don’t ship to customers — instead, the bot points to the closest related pages, and visitors find their answer 71% of the time without opening a ticket.
    • Plug into your existing helpdesk. Native handoff for Zendesk, Intercom, HubSpot, Front, and Help Scout. The bot handles tier-1, your team handles the rest with full conversation context.
    • Theme to match your brand in eight CSS properties. Three preset themes (FlowChat default, Stripe-like, Linear-like) cover most cases at one click.
    • Five pricing tiers from $49/mo Starter to custom Enterprise, with overage at $0.005/answer above plan — so a viral support week is a $250 line item, not a contract renegotiation.

    Known caveats for v0.1:

    • SAML/OIDC enterprise auth is in private beta — request access via enterprise@flowchat.com.
    • SOC 2 Type II is in active audit (observation period started Q2 2026). Type I report under NDA.
    • Multi-locale crawling is limited to UTF-8 sites in this release; CJK chunking lands in 0.2.
    Under the hood (for the engineering-curious)

    For teams who want to know how the answer-quality machinery actually works:

    • Crawler: Cloudflare Browser Rendering for JS-aware fetching, robots.txt + Content Signals compliance per request, conditional GET dedup with ETag/Last-Modified tracking. The Crawl Compliance Inspector on /product/crawl walks through the per-URL admission policy.
    • Chunker: heading-aware semantic splits with table/code-block preservation, late chunking on bge-m3’s 60K-token context window, Anthropic-style contextual retrieval prefixes added at ingest.
    • Retrieval: Vectorize for dense + D1 FTS5 for BM25, fused with Reciprocal Rank Fusion (default k=60), reranked by Cohere Rerank 3.5 to top-3.
    • Generation: Llama 3.3 70B (default) or Claude Sonnet 4.6 via AI Gateway (premium routing on complex intents). Token-level [#N] citation discipline; output without citations is regenerated.
    • NLI verification: every cited claim is checked for entailment by a RAG-fine-tuned NLI model. Sub-threshold claims are stripped post-hoc.
    • Refusal gate: when reranker top-3 score is below threshold, generation doesn’t run. The widget renders refusal in a distinct purple, with the closest related pages.

    The full architecture is on /developers and /security.