All projects
AI / MLResearch2026
No Transformer
Retrieval engine without transformer models
Role / Solo — research

- transformer models
- 0
- pytest suite
- Tested
- local knowledge base
- Offline
The problem
Transformer-based RAG is expensive to run and hard to debug. For narrow domains, classical retrieval can be cheaper and fully explainable.
Approach
- Built modular retrieval engines with a shared interface so strategies can be swapped.
- Kept a local knowledge base and scored answers without embedding models.
- Covered the pipeline with pytest and kept the whole thing runnable offline.
Result
- Working retrieval pipeline that runs without GPU or API cost.
- Every answer step can be traced back to a document and a scoring rule.
Stack
- Python
- FastAPI
- pytest
- Custom engines