How raywincruz07-collab/context-matters-diversity-aware-rag works
This is a reproducibility package for research evaluating whether diverse retrieval contexts improve RAG answers across PubMedQA, HotpotQA, and ASQA. Its principal workflow retrieves Top-20 candidates, selects a fixed Top-5 context using a diversification condition, generates an answer, then evaluates it in Phase 4 and analyzes cross-metric relationships in Phase 5. The graph also includes provenance/artifact and metric-contract layers. Several experimental subsystems are represented at their documented directory level because their implementation was not sampled; those links rely on README statements rather than inspected runtime calls.
Main entry points: diversification Diversification · producer.py Candidate artifacts · evaluation Phase 4 metrics · selected_context.py Top-5 context
Components 20
Retrieval and provenance
- Retriever families
sprint1_baseline - Candidate artifacts
sprint1_baseline/asqa/src/retrieval_artifacts/producer.py - Artifact contracts
sprint1_baseline/asqa/src/retrieval_artifacts/contracts.py - Corpus manifests
sprint1_baseline/asqa/src/retrieval_artifacts/runtime_corpus.py - Retriever configs
sprint1_baseline/asqa/src/retrievers
Context selection
- Diversification
sprint2_diversification/src/diversification - MMR selection
sprint2_diversification/src/diversification/mmr.py - DPP selection
sprint2_diversification/src/diversification/dpp.py - Cluster selection
sprint2_diversification/src/diversification/clustering.py - Top-5 context
sprint1_baseline/hotpotqa/src/generation/selected_context.py
Answer generation
- LLM generation
sprint2_diversification/src/generation - Generation runtime
sprint1_baseline/asqa/src/generation/maki.py - Artifact I/O
sprint1_baseline/asqa/src/generation/_io.py - Generated answer
Evaluation and analysis
- Phase 4 metrics
sprint3_evaluation/src/evaluation - Metric registry
sprint3_evaluation/src/evaluation/metric_registry.py - Metric contracts
sprint3_evaluation/src/evaluation/contracts.py - Phase 5 analysis
sprint3_evaluation/src/evaluation
Experiment governance
- Run provenance
sprint1_baseline/asqa/src/run_registry.py
Other components
- Question
Connections 19
- Question → Retriever families: queries
README.md - Retriever families → Candidate artifacts: produce candidates
README.md - Candidate artifacts → Diversification: candidate pool
README.md - Diversification → Top-5 context: selects Top-5
README.md - Diversification → MMR selection: dispatches
- Diversification → DPP selection: dispatches
sprint2_diversification/src/diversification/dpp.py - Diversification → Cluster selection: dispatches
- Top-5 context → LLM generation: supplies context
README.md - LLM generation → Generated answer: generates
README.md - Generation runtime → Artifact I/O: uses
sprint1_baseline/asqa/src/generation/maki.py - LLM generation → Generation runtime: runs through
sprint2_diversification/src/generation/cli_support.py - Generated answer → Phase 4 metrics: evaluated in
README.md - Top-5 context → Phase 4 metrics: evaluates evidence
README.md - Phase 4 metrics → Phase 5 analysis: cross-metric analysis
README.md - Phase 4 metrics → Metric registry: checks applicability
sprint3_evaluation/src/evaluation/query_evaluator.py - Phase 4 metrics → Metric contracts: validates results
sprint3_evaluation/src/evaluation/query_evaluator.py - Candidate artifacts → Artifact contracts: validates artifacts
sprint1_baseline/asqa/src/retrieval_artifacts/producer.py - Corpus manifests → Candidate artifacts: uses producers
sprint1_baseline/asqa/src/retrieval_artifacts/runtime_corpus.py - Candidate artifacts → Retriever configs: records provenance
sprint1_baseline/asqa/src/retrieval_artifacts/producer.py