How optimal-cognitive-core/occ-rag works
OCC-RAG is a release of compact models for context-grounded question answering, not an implemented RAG pipeline in this repository. The documented workflow supplies a question and source documents, formats them with the model’s chat template, generates a structured response, and reads its answer or abstention. The README documents Transformers and vLLM usage; the examples are present in the tree but were not sampled, so implementation-level connections are limited to those documented relationships.
Main entry points: README.md Structured prompt · quickstart_transformers.py Transformers quickstart · quickstart_vllm.py vLLM quickstart
Components 6
Prompt and response format
- Structured prompt
README.md - Five-section response
README.md
Model inference examples
- Transformers quickstart
examples/quickstart_transformers.py - vLLM quickstart
examples/quickstart_vllm.py
Other components
- Questioner
- OCC-RAG models
Connections 6
- Questioner → Structured prompt: provides question and sources
README.md - Structured prompt → OCC-RAG models: formats model input
README.md - OCC-RAG models → Five-section response: generates structured trace
README.md - Transformers quickstart → OCC-RAG models: loads and generates
README.md - vLLM quickstart → OCC-RAG models: serves inference
README.md - Five-section response → Questioner: returns answer or abstention
README.md