How majdzidan-dev/agentic-rag works
AgenticRag is a command-line research assistant. The user enters a question in the chat loop; the agent selects a calculator, general web search, or PubMed research tool, then returns a response with scientific citations where applicable. PubMed search checks local coverage, fetches and chunks papers when needed, indexes them, retrieves and reranks passages, and can grade and rewrite queries before returning excerpts. The README documents an isolated retrieval-evaluation capability as well. Edges reflect the excerpts, SOURCE INDEX references, and README; unsampled implementation details are not inferred.
Main entry points: pubmed_rag.py PubMed Search Tool · agent.py Gemini Agent · run_eval.py Evaluation Runner · retrieval_grader.py Retrieval Grader
Components 20
Conversation
- Chat Loop
main.py - Gemini Agent
agent.py - API Configuration
config.py
Assistant Tools
- Calculator
tools/calculator.py - Web Search
tools/web_search.py
PubMed Research
- PubMed Search Tool
tools/pubmed_rag.py - Abstract Chunking
tools/pubmed_rag.py - Hybrid Search
rag/hybrid_search.py - Cross-Encoder Reranker
rag/reranker.py - Retrieval Grader
rag/retrieval_grader.py - LLM Text Helper
llm_utils.py
Knowledge Stores
- Vector Store
rag/vector_store.py - BM25 Index
rag/keyword_store.py
Retrieval Evaluation
- Evaluation Runner
eval/run_eval.py - Retrieval Variants
eval/retrieval_variants.py - Ranking Metrics
eval/metrics.py
Other components
- User
- Gemini
- SerpAPI
- PubMed API
Connections 31
- User → Chat Loop: enters question
README.md - Chat Loop → Gemini Agent: invokes
main.py - Gemini Agent → Gemini: uses model
agent.py - Gemini Agent → Calculator: dispatches
agent.py - Gemini Agent → Web Search: dispatches
agent.py - Gemini Agent → PubMed Search Tool: dispatches
agent.py - Web Search → SerpAPI: searches
tools/web_search.py - PubMed Search Tool → PubMed API: fetches papers
tools/pubmed_rag.py - PubMed Search Tool → Abstract Chunking: chunks abstracts
tools/pubmed_rag.py - Abstract Chunking → Vector Store: indexes chunks
tools/pubmed_rag.py - Abstract Chunking → BM25 Index: indexes chunks
tools/pubmed_rag.py - PubMed Search Tool → Vector Store: checks coverage
tools/pubmed_rag.py - PubMed Search Tool → Hybrid Search: retrieves candidates
tools/pubmed_rag.py - Hybrid Search → Vector Store: queries dense index
rag/hybrid_search.py - Hybrid Search → BM25 Index: queries BM25
rag/hybrid_search.py - PubMed Search Tool → Cross-Encoder Reranker: reranks candidates
tools/pubmed_rag.py - PubMed Search Tool → Retrieval Grader: grades excerpts
tools/pubmed_rag.py - Retrieval Grader → Gemini: grades and rewrites
rag/retrieval_grader.py - Retrieval Grader → LLM Text Helper: extracts response text
rag/retrieval_grader.py - API Configuration → Gemini Agent: provides API key
agent.py - API Configuration → Web Search: provides API key
tools/web_search.py - API Configuration → Retrieval Grader: provides API key
rag/retrieval_grader.py - API Configuration → Vector Store: provides API key
rag/vector_store.py - Evaluation Runner → Vector Store: resets eval collection
eval/run_eval.py - Evaluation Runner → BM25 Index: resets eval corpus
eval/run_eval.py - Evaluation Runner → PubMed Search Tool: indexes known papers
eval/run_eval.py - Evaluation Runner → Retrieval Variants: runs variants
eval/run_eval.py - Retrieval Variants → Hybrid Search: runs hybrid search
eval/retrieval_variants.py - Retrieval Variants → Cross-Encoder Reranker: reranks candidates
eval/retrieval_variants.py - Retrieval Variants → Retrieval Grader: grades and rewrites
eval/retrieval_variants.py - Evaluation Runner → Ranking Metrics: scores rankings
eval/run_eval.py