RPRajat Pal
← All work

Infosys · for a US bank

Batch RAG Ingestion Pipeline

An end-to-end pipeline that turns raw documents in S3 into a searchable vector index, tuned for accurate, context-aware retrieval.

  • Python
  • Amazon S3
  • Embeddings
  • Vector store
  • Chunking
  • Metadata filtering

The problem

Several GenAI use cases needed to answer questions from large, changing document sets. Answer quality depended mostly on retrieval: when the wrong chunks came back, even a strong model gave weak answers.

What I built

A batch RAG pipeline shared across use cases:

  1. Ingestion from Amazon S3.
  2. Document chunking, with the strategy chosen per document type.
  3. Embedding generation for every chunk.
  4. Vector store indexing, with metadata kept alongside each vector.

Improving retrieval

The biggest gains came from the retrieval side, not the model:

  • Better chunking. Chunk size and boundaries decide whether a retrieved passage holds a complete idea. Tuning them gave more relevant context.
  • Metadata filtering. Narrowing the search by document attributes before semantic search removes results that look similar but are irrelevant.

The result was more accurate, context-aware responses across multiple GenAI use cases. These changes were validated with the evaluation practices I set up.