Skip to content
← all systems

SignalDesk

RAG support chat that answers only from your indexed knowledge base - and says so when it cannot.

grounded answers · zero hallucinated claims by design

role
Architect and sole engineer
stack
Next.js 16 · React 19 · TypeScript · Supabase · OpenAI · RAG
status
active

SignalDesk is a retrieval-augmented support chat that answers only from an indexed knowledge base. It ingests technical docs, API references, and product content, and constrains every answer to the sources it retrieves - so the bot either answers from the docs and cites where the answer came from, or explicitly says it cannot when retrieval confidence is low. The refusal path is a designed feature, because a support bot that guesses about pricing or an API destroys trust faster than no bot at all.

// 01 - PROBLEM

Support chatbots that improvise are worse than no chatbot: a wrong answer about pricing or an API costs trust. SignalDesk indexes a product knowledge base and constrains answers to retrieved sources, so the bot either answers from the docs or admits it cannot.

// 02 - APPROACH

  • Knowledge-base ingestion and indexing for technical docs, API references, and product content.
  • Retrieval-constrained generation: answers cite the indexed source they came from.
  • Explicit refusal path when retrieval confidence is low - honesty over fluency.

// 03 - ARCHITECTURE

▸ retrieval-grounded · refuses when unsure

  1. Docs

    Technical docs, API references, product content

  2. Index

    Ingest and index the knowledge base

  3. Retrieve

    Fetch the sources relevant to the question

  4. Generate

    Answer constrained to sources; refuse when confidence is low

  5. Cite

    Answer with the indexed source it came from

Refusal as a feature
A support bot that says "I do not have that in the docs" preserves trust; one that guesses destroys it. The refusal path is designed, not an accident.

// 04 - PRODUCTION-GRADE

  • Live deployment at signaldesk.mikul.me
  • Retrieval grounding with source citation
  • Analytics on demo access and query patterns

// 05 - ARTIFACTS