Fintech Ticket Triage with Node.js RAG: PDF Upload, Semantic Search, and Citations

To handle PDF uploads and perform semantic search using Node.js RAG (Retrieval-Augmented Generation), the process starts with parsing the uploaded PDF and dividing it into overlapping chunks. Each chunk is then embedded with its filename, page, and section information. These embedded chunks are subsequently retrieved using pgvector, a metadata-aware semantic search tool.

The retrieved chunks and their citation metadata are then passed to the answer generation model. This pipeline ensures that the generated answers are contextually relevant and properly cited. The retrieval result carries forward the evidence, eliminating the need for the answer model to recreate provenance from memory. This approach is particularly useful in a fintech support queue, where structured output correctness is the release gate, rather than dashboard scores or fluent demos.

It ensures that the relevant page is identified, and the on-call engineer can reproduce the answer using the cited source. The key to a successful implementation is determining the optimal chunk size, which should balance between being large enough to include necessary context and small enough to avoid separating conditions from their exceptions.

Overlap between chunks helps maintain relationships between related information. Token counting is essential to ensure that the chunk size and top-k context fit within the answer prompt. The final output includes a binary check for malformed responses, missing citation metadata, and citations that do not resolve to retrieved chunk IDs.

This binary check distinguishes between bad model responses and stale retrieval indexes, providing valuable feedback to the responder about where to start troubleshooting.