Signal Support Copilot
A retrieval-grounded assistant that drafts support replies from a company's own knowledge.
Support volume was growing faster than the team could hire. Off-the-shelf chatbots either hallucinated or deflected everything to a generic FAQ, and agents didn't trust them.
We built a retrieval pipeline over the company's own content with strict grounding: the assistant answers only from retrieved passages, always cites them, and routes low-confidence questions to a human agent together with the supporting material it has already gathered.
Grounded drafting
Replies are generated only from retrieved, cited source passages.
Confidence gating
Low-confidence questions are escalated with context instead of guessed at.
Answer review loop
Agent edits feed back into evaluation so quality is measured, not assumed.
Knowledge gaps report
Surfaces the questions the knowledge base can't yet answer.
- 01Frontend
Agent console embedded in the existing helpdesk via extension
- 02API Layer
Streaming completion API with per-tenant rate limits and PII redaction
- 03Services
Ingestion, embedding, retrieval and a LangGraph orchestration layer
- 04Database
Postgres with pgvector for embeddings; document store for source content
- 05External Integrations
Helpdesk, knowledge base and model providers
Frontend
- Next.js
- TypeScript
- Tailwind CSS
Backend
- Python
- FastAPI
- LangGraph
Database
- PostgreSQL
- pgvector
Cloud
- Azure Container Apps
- Blob Storage
AI
- Claude
- OpenAI embeddings
- Ragas evals
DevOps
- Bicep
- GitHub Actions
- Langfuse
- Every answer traceable to a source
- Low-confidence questions escalated rather than answered speculatively
- Knowledge gaps made visible to the content team
- Evaluation built in from the outset