DocsChat

Businesses publish thousands of pages of documentation — help centres, product pages, onboarding guides, FAQs — but customers still can't find answers quickly. They open support tickets, abandon products, or leave. The knowledge is there; it just isn't accessible.

There was no turnkey solution that:

• Let a non-technical business owner point at a URL and get a live AI chat widget back
• Required zero ML knowledge to configure
• Produced a two-line embed snippet compatible with any website

DocsChat is a multi-tenant SaaS platform that turns any URL or document library into a context-aware AI chat widget — in under three minutes.

A user pastes a URL, the platform crawls and indexes the content automatically, and they get a snippet they drop into any site. No config files. No API wrangling. No AI expertise needed.

URL / Docs  →  Crawl  →  Chunk  →  Embed  →  pgvector  →  Claude Answer

• Crawl — automated spider discovers up to 100 pages from a seed URL, respecting robots.txt
• Chunk — content split into overlapping 512-token segments to preserve context
• Embed — each chunk encoded to a 768-dimension vector via Google Gemini text-embedding-004
• Store — vectors persisted in Supabase pgvector with cosine similarity indexing
• Answer — top-8 chunks retrieved at query time; Anthropic Claude composes a cited, grounded answer

For business owners

• No-code setup — paste a URL, wait ~2 minutes, get a widget
• Live ingestion progress — real-time SSE stream shows crawl and indexing phases
• Re-index on demand — one click refreshes the knowledge base from source
• Conversation history — all chats stored and browsable per knowledge base
• API key management — scoped keys for widget access, SHA-256 hashed at rest

For developers

• Iframe-isolated widget — zero CSS bleed, zero dependency, sandboxed
• Streaming responses — Claude answers stream token-by-token
• Chunk browser — inspect every stored chunk, word counts, and metadata
• REST API — full programmatic access to knowledge bases and chat

For agencies

• Multi-tenant — unlimited knowledge bases per account, fully isolated via RLS
• White-label ready — brand colour, position, and welcome message configurable per widget
• Anonymous trial flow — landing page lets visitors chat against any URL with no sign-up

RAG pipeline with zero cold-start

Ingestion runs as a FastAPI BackgroundTask. Progress is written to an ingest_progress JSONB column in Supabase every few seconds. The frontend opens an SSE connection and streams updates live — no polling, no WebSocket overhead.

Free embeddings at scale

Google Gemini text-embedding-004 produces 768-dimension vectors on the free tier (1,500 requests/minute). Switching from OpenAI embeddings cut embedding cost to zero without sacrificing retrieval quality.

Widget CSS isolation

The embed is a sandboxed <iframe> rather than injected DOM elements. Host-page styles can't bleed in; widget styles can't leak out. Communication between host and widget happens via postMessage.

Multi-tenant security

Every database table has Row Level Security enabled. The backend uses Supabase's service role key only for background ingestion; all user-facing reads use the anon key + JWT, so users can only ever access their own data.

• Enables instant self-service answers, reducing support ticket volume
• No-code setup means sales cycle is the live demo — visitors try it on the landing page before signing up
• Free embedding tier keeps margins high at small-to-medium scale
• Widget works on WordPress, Webflow, Shopify, or custom stacks — no integration work needed
• Resellable as a managed service: one-time setup fee + monthly retainer per client