sachhh05/adaptive_rag

How sachhh05/adaptive_rag works

This is an experimental adaptive RAG system: queries are routed to direct answering, retrieval, or decomposition, then answers and evidence are scored and results recorded. The README describes LinUCB routing, while the sampled evaluation excerpt also imports a rule-based router and LinUCB components; the omitted evaluation code prevents confirming the exact policy dispatch and update sequence. Edges reflect documented behavior or visible source references, with uncertain wiring left out.

Main entry points: checkpoint5_full_eval.py CP5 Evaluator · reward_function.py Reward Composer · retrieve.py Retrieve Action · decompose.py Query Decomposition

Components 20

Query Routing

Query Features pipeline/feature_extractor.py
Rule Router pipeline/adaptive_router.py
Context Builder pipeline/bandit/context.py
LinUCB Policy pipeline/bandit/linucb.py

Answer Actions

Direct Answer pipeline/actions/answer_directly.py
Retrieve Action pipeline/actions/retrieve.py
Query Decomposition pipeline/actions/decompose.py
Reranking pipeline/actions/rerank.py
Action Utilities pipeline/actions/common.py

Scoring and Evaluation

CP5 Evaluator eval/checkpoint5_full_eval.py
Reward Composer pipeline/reward/reward_function.py
Answer Quality pipeline/reward/answer_quality.py
Faithfulness Scorer pipeline/reward/faithfulness.py
Retrieval Precision pipeline/reward/retrieval_precision.py
Latency Scorer pipeline/reward/latency.py
Evaluation Results results/checkpoint5/checkpoint5_results.csv

Knowledge and Baselines

ChromaDB Collection
Groq LLM API
Naive RAG Baseline pipeline/naive_rag.py

Other components

Questioner

Connections 25

Questioner → CP5 Evaluator: submits query README.md
CP5 Evaluator → Query Features: extracts features eval/checkpoint5_full_eval.py
Rule Router → Query Features: extracts features pipeline/adaptive_router.py
CP5 Evaluator → Context Builder: builds context eval/checkpoint5_full_eval.py
CP5 Evaluator → LinUCB Policy: uses policy eval/checkpoint5_full_eval.py
Query Features → Context Builder: normalizes features README.md
Context Builder → LinUCB Policy: supplies context README.md
CP5 Evaluator → Direct Answer: executes action README.md
CP5 Evaluator → Retrieve Action: executes action README.md
CP5 Evaluator → Query Decomposition: executes action README.md
Query Decomposition → Retrieve Action: retrieves subquestions pipeline/actions/decompose.py
Retrieve Action → Action Utilities: uses utilities pipeline/actions/retrieve.py
Direct Answer → Action Utilities: uses utilities pipeline/actions/answer_directly.py
Retrieve Action → ChromaDB Collection: queries chunks pipeline/actions/retrieve.py
Retrieve Action → Groq LLM API: generates answer
Query Decomposition → Groq LLM API: generates subquestions pipeline/actions/decompose.py
Query Decomposition → Groq LLM API: synthesizes answer pipeline/actions/decompose.py
Naive RAG Baseline → ChromaDB Collection: retrieves chunks pipeline/naive_rag.py
Naive RAG Baseline → Groq LLM API: generates answer pipeline/naive_rag.py
CP5 Evaluator → Reward Composer: computes reward eval/checkpoint5_full_eval.py
Reward Composer → Answer Quality: scores answer pipeline/reward/reward_function.py
Reward Composer → Faithfulness Scorer: scores support pipeline/reward/reward_function.py
Reward Composer → Retrieval Precision: scores evidence pipeline/reward/reward_function.py
Reward Composer → Latency Scorer: scores latency pipeline/reward/reward_function.py
CP5 Evaluator → Evaluation Results: writes results eval/checkpoint5_full_eval.py

Drawn by GitDiagram on November 11, 2026 from the repository's file tree, README and source files. Read as Markdown ·