How jayeshvadje/telecom-support-intelligence works
This repository implements a telecom-policy RAG service plus a document-indexing workflow. Policy documents are parsed, chunked, enriched, embedded, and upserted to Pinecone; API queries are embedded and searched, then Gemini generates grounded answers with source metadata. A separate evaluator scores answers. The source excerpts establish these runtime calls, but no README was provided and app.py was not excerpted, so its role is not represented.
Main entry points: pipeline.py Ingestion pipeline · rag_chain.py RAG answer chain · pinecone_retriever.py Semantic retriever · evaluator.py RAG evaluator
Components 15
Policy ingestion
- Policy documents
ingestion/data - Ingestion pipeline
ingestion/pipeline.py - DOCX loader
ingestion/loaders/docx_loader.py - Text cleaner
ingestion/processors/text_cleaner.py - Policy chunker
ingestion/chunking/policy_chunker.py - Metadata extractor
ingestion/metadata/extractor.py
Vector indexing
- Document embedder
ingestion/embeddings/gemini_embedder.py - Index manager
ingestion/vectorstore/pinecone_manager.py - Pinecone index
Query service
- Query API
main.py - RAG answer chain
generation/rag_chain.py - Semantic retriever
retrieval/pinecone_retriever.py
Answer evaluation
- RAG evaluator
evaluation/evaluator.py
Other components
- Support user
- Google Gemini
Connections 18
- Policy documents → Ingestion pipeline: provides files
ingestion/pipeline.py - Ingestion pipeline → DOCX loader: parses documents
ingestion/pipeline.py - DOCX loader → Text cleaner: cleans text
ingestion/loaders/docx_loader.py - Ingestion pipeline → Policy chunker: creates chunks
ingestion/pipeline.py - Ingestion pipeline → Metadata extractor: enriches chunks
ingestion/pipeline.py - Ingestion pipeline → Document embedder: embeds chunks
ingestion/pipeline.py - Document embedder → Google Gemini: requests embeddings
ingestion/embeddings/gemini_embedder.py - Ingestion pipeline → Index manager: upserts chunks
ingestion/pipeline.py - Index manager → Pinecone index: writes vectors
ingestion/vectorstore/pinecone_manager.py - Support user → Query API: submits query
main.py - Query API → RAG answer chain: requests answer
main.py - RAG answer chain → Semantic retriever: searches policies
generation/rag_chain.py - Semantic retriever → Google Gemini: embeds query
retrieval/pinecone_retriever.py - Semantic retriever → Pinecone index: queries vectors
retrieval/pinecone_retriever.py - RAG answer chain → Google Gemini: generates answer
generation/rag_chain.py - RAG evaluator → RAG answer chain: gets answer
evaluation/evaluator.py - RAG evaluator → Semantic retriever: gets context
evaluation/evaluator.py - RAG evaluator → Google Gemini: scores response
evaluation/evaluator.py