EXP-003 / question
Technical question
Can retrieval paths make RAG answers more explainable, grounded, and easier to inspect?
EXP-003 / method
Method and workflow
- Load and split documents into retrievable chunks.
- Build embeddings and FAISS-based semantic search.
- Construct concept and chunk relationships in an inspectable graph layer.
- Expand retrieval results with graph neighbors and ranking logic.
- Return extractive or optional LLM answers with source citations.
- Evaluate retrieval behavior using golden queries and automated tests.
documents
chunks
embeddings
knowledge graph
hybrid retriever
citations
answer UI
EXP-003 / evidence
Evidence of work
Hybrid retrieval
Vector search and graph expansion work together to improve context discovery and inspectability.
Evaluation
Golden-query metrics make retrieval quality measurable instead of anecdotal.
Interface
FastAPI and Streamlit layers make the system usable as a small RAG application.
EXP-003 / stack
Technical stack
PythonFastAPIPydanticUvicornSentence TransformersFAISSNetworkXNumPyStreamlitunittestGitHub Actions
Open repository ↗
EXP-003 / limitations
Limitations and honesty check
- The graph layer is in-memory and would need persistence such as Neo4j for larger deployments.
- Answer quality still depends on document coverage and retrieval quality.
- Production use would need authentication, monitoring, cost controls, and citation validation.
EXP-003 / next
Next improvements
- Add persistent graph storage.
- Add larger benchmark corpora and citation-accuracy tests.
- Add Docker deployment and API documentation.
- Add feedback-based retrieval tuning.