Skip to main content
All work

AI · RAG · Next.js · 2025

DocPulse

Document Q&A for a legal-tech startup.

Next.jspgvectorOpenAIHybrid retrievalEval harnessTypeScript

Client type

AI

Stack

Next.js · pgvector · OpenAI · Hybrid retrieval

Timeline

2025

Outcome

Citation accuracy on the eval set rose from about 60% with prompt-only tuning to 94% after retrie…

Case study

Problem

Problem

The team pasted contracts into ChatGPT and got hallucinated clause references. They needed grounded answers with page citations and a way to re-index when documents changed.

Solution

Solution

We built a RAG pipeline with PDF ingestion, hybrid vector and keyword retrieval, citation rendering, and a 20-question eval set that runs before every prompt change.

Implementation

Implementation

  • PDF ingestion pipeline with chunking and metadata extraction
  • pgvector store with hybrid retrieval and reranking
  • Chat UI with inline citations linked to source pages
  • Admin panel to upload, re-index, and review low-confidence chunks
  • Eval harness with golden questions and regression alerts

Demo

Demo

demo · docpulse

>Want one like this?

Open

Outcome

Outcome

Citation accuracy on the eval set rose from about 60% with prompt-only tuning to 94% after retrieval work. Legal review time per contract dropped from 45 minutes to under 10.

Engineering insight

Engineering insight

If you are building RAG, write evals before you tune prompts. Retrieval quality drove results more than model choice. Measure citations, not vibes.

Want something like DocPulse? Tell us on the intake form and we will scope your version.

Start a project