← Back to projects

Case study

Legal/Medical Research Engine

Cited extractive answers over a large EU legal and medical corpus without storing queries.

At a glance

Plain summary for recruiters and visitors. Technical detail follows below.

What it is
I built a research search tool that answers questions over EU legal and medical documents with cited passages. Queries are not stored by design.
What I owned
Solo end to end. Ingestion, chunking, embedding index, retrieval, answer assembly, privacy design, and demo deployment.
Why it matters
Analysts and researchers save time when they need a cited passage instead of reading long PDF sets manually. Confidence scores show when retrieval is weak.
Try it on this page vs full project
The try-it box searches a small curated corpus on Cloudflare Workers. The LLMOps platform repo with Helm, GitOps, and eval CI is on request. The full document index with Chroma and FastAPI is also on request.

Quick comparison

  • This page demo

    Curated Worker corpus. Hybrid retrieval and extractive answers with citations. No query storage.

  • Production stack (on request)

    Helm on k3d, Argo CD GitOps, golden eval CI, Prometheus metrics, RUNBOOK, and AWS deploy (ECR plus EC2 free tier or App Runner in eu-central-1). Repo access for hiring managers.

  • Full stack on request

    Full Chroma index, larger document set, and FastAPI backend from the private repo.

Legal/Medical Research Engine

Problem

Legal and medical professionals need fast answers over large document sets without storing sensitive query content.

Method

Built a RAG pipeline over 65,000+ documents with vector retrieval, citations, and confidence scores. Designed so queries are not stored.

Result

Delivers cited answers quickly for EU legal and medical guidelines without storing queries. Recruiters can try a curated query on this page. The full 65,000 document index stays on request.

Architecture

How the system is shaped. Full implementation stays private.

  1. Step 1

    Ingest

    Chunk EU legal and medical guideline text with provenance metadata.

  2. Step 2

    Index

    Embedding model + vector store (full Chroma stack in private repo). Free demo uses a curated Worker corpus.

  3. Step 3

    Retrieve

    Hybrid semantic + keyword retrieval with top-k passages.

  4. Step 4

    Answer

    Extractive synthesis with citations and confidence. No paid LLM required for the portfolio demo.

  5. Step 5

    Privacy

    Queries are not persisted. Portfolio demo discloses curated-subset limits.

  6. Step 6

    Platform

    Helm chart, k3d lab cluster, Argo CD GitOps, GitHub Actions eval CI, Prometheus metrics, and optional AWS path (ECR image, EC2 free tier or App Runner) in the private platform repo.

Public demo scope

The in-page widget uses a curated Worker corpus. Queries are not stored. The full 65,000 document index stays on request.

ScopeStatusNote
Public demo corpusCurated subset
Query storageNone
Full document indexOn request

Algorithm

Hybrid retrieve-then-extract

Score passages by query overlap / semantic similarity, then build a cited answer from top excerpts.

tokens = tokenize(question)
candidates = rank(corpus, tokens, top_k)
answer = stitch_excerpts(candidates)  # with [1][2] citations
confidence = f(top_score, k)
return answer, sources, confidence

Key logic

Compact illustrative snippet (Hybrid retrieval score (illustrative)). Not the full codebase.

function hybridScore(doc, queryTokens, queryVec, alpha = 0.65) {
  const bm25 = bm25Score(doc.tokens, queryTokens);
  const semantic = cosine(queryVec, doc.embedding);
  return alpha * semantic + (1 - alpha) * bm25;
}

function rankPassages(corpus, question, topK = 8) {
  const tokens = tokenize(question);
  const queryVec = embed(question);
  return corpus
    .map((doc) => ({ doc, score: hybridScore(doc, tokens, queryVec) }))
    .sort((a, b) => b.score - a.score)
    .slice(0, topK);
}

Stack

RAGAIVector DatabasesDocument SearchNatural Language ProcessingPrivacy-by-Design

Try a live query

Free Cloudflare demo over a curated corpus. Not the full private index. Queries are not stored on the server. Recent tries stay in your browser only.

Suggested (click to run):

Source code

Full source is available to hiring managers on request. The public page shows architecture, algorithms, and compact proofs only.