How it’s built
A small local model, and the system that keeps it honest.
One press
Code picks the specialists a question needs. A guard checks every quote.
Contradictions, checked in code
A fast checker first. A second look only when it is unsure.
- F1 on DECODE
- 0.913
- 0.753 for the baseline
- False alerts
- Halved
- Against the baseline
- Time per check
- 45 ms
- On 1,000 published pairs
Finding the right turn
Keywords and meaning, then a reranker.
The stack
- Speech
- whisper.cpp in C++ behind gRPC, with pyannote speaker labels.
- speech/
- Backend
- FastAPI, one WebSocket per meeting.
- src/contextlens/api/
- Memory
- PostgreSQL with pgvector.
- src/contextlens/
- Retrieval
- BM25, embeddings, cross-encoder reranker.
- src/contextlens/retrieval/
- Agents
- LangGraph. Code routes to the specialists.
- src/contextlens/agents/
- Model
- A 4B model through Ollama.
- qwen3.5:4b-q4_K_M
- Clients
- Desktop, Unity for glasses, and a web client.
- clients/
Evaluation
161 held-out cases from QMSum and AMI, with conflicts 20 to 60 turns apart. The shipped system against the original single agent.
| Measure | Single agent | ContextLens Neo |
|---|---|---|
| Conflict catch rate | 0% | 87% |
| Found evidence 10+ turns back | 0% | 60% |
| Answers graded correct | 35% | 65% |
| Tokens per card | 711 | 630 |
| Median latency, ms | 2,044 | 2,336 |