AI Research System with Verified Knowledge Retrieval
Medical AI that cites every claim — no hallucinations, no guessing.
MediCortex answers medical questions by searching 50,000+ peer-reviewed PubMed papers in real time and grounding every response in what the research actually says. Every fact links to the source paper. Every answer is reproducible.
🌐 Try It Live | 🎥 Watch Demo
General-purpose AI models are trained on internet text — which includes misinformation, outdated advice, and content that confidently states things that are simply wrong. In healthcare, that is not acceptable.
MediCortex takes a different approach: the AI never generates claims it cannot back up with a source. If the research doesn't support it, the answer says so.
MediCortex implements a RAG (Retrieval-Augmented Generation) pipeline purpose-built for scientific literature:
User Question
│
▼
Vertex AI text-embedding-004
│ 768-dimensional query vector
▼
Qdrant Vector Search ←──────── 50,000+ indexed PubMed abstracts
│ Top-k semantically similar papers
▼
Context Assembly
│ Title · Abstract · PMID · Authors · Year
▼
Google Gemini (grounded prompt — citation required for every claim)
│
▼
Answer with [PMID: xxx] citations → verifiable on PubMed
The constraint is the point. Gemini is instructed it may only use information from the retrieved papers. If the top results are weak, the answer reflects that uncertainty rather than fabricating a confident response.
| Feature | Description |
|---|---|
| Cited Q&A | Every factual claim links to its PubMed source — click any [PMID] to verify |
| Semantic Search | 768-dim vector embeddings find conceptually related papers, not just keyword matches |
| Multi-Paper Synthesis | Select several papers and get a structured analysis of where researchers agree and disagree |
| Study Comparison | Side-by-side breakdown of methodology, sample size, findings, and limitations |
| Key Insight Extraction | Distill the most actionable clinical findings from a set of papers |
| Cross-Platform | Flutter — runs on web, macOS, Windows, and Linux from a single codebase |
Google Gemini — Response generation with a grounding constraint system prompt that forces citation of all claims and flags when evidence is insufficient.
Vertex AI text-embedding-004 — Converts both queries and documents into 768-dimensional semantic vectors. Uses separate task types (RETRIEVAL_QUERY vs RETRIEVAL_DOCUMENT) for optimized similarity scoring.
Qdrant — Vector database hosting 50,000+ article embeddings. Handles cosine similarity search for vector queries and payload text-index scroll for keyword fallback when embeddings are unavailable.
// 1. Embed the query
final embedding = await vertexAI.generateQueryEmbedding(question);
// 2. Retrieve the most relevant papers
final papers = await qdrant.hybridSearch(
query: question,
queryEmbedding: embedding,
size: 5,
);
// 3. Build grounded context
final context = buildContext(papers); // title, abstract, PMID, year
// 4. Generate with citation constraint
final answer = await gemini.generate(
systemPrompt: "Answer ONLY using the papers below. Cite [PMID: x] for every claim.",
context: context,
question: question,
);- Source: PubMed via NCBI E-utilities (free, public API)
- Scale: 50,000+ articles across 100+ medical topics
- Fields: PMID, title, abstract, authors, publication date, journal, article type
- Vectors: 768-dim Vertex AI embeddings, indexed in Qdrant with cosine similarity
- Coverage: Chronic disease, oncology, mental health, neurology, infectious disease, autoimmune, respiratory, nutrition, preventive care, pediatrics, women's health, geriatrics, precision medicine, emerging therapies
| Layer | Technology |
|---|---|
| UI | Flutter · Material Design 3 · Riverpod |
| AI | Google Gemini API · Vertex AI Embeddings |
| Search | Qdrant Cloud (vector DB) |
| Proxy | Appwrite Functions (CORS handling for web) |
| Ingestion | Python · NCBI E-utilities · google-auth |
| Architecture | Clean Architecture · Repository pattern |
- Flutter SDK
- Google Cloud project with Gemini API and Vertex AI enabled
- Qdrant Cloud account (free tier supports this dataset)
- Appwrite account (for web deployment proxy)
git clone https://github.com/mj-963/medicortex.git
cd medicortex
cp env.example.json env.json
# Fill in: qdrant, vertex_ai, service_account, gemini_api_keyflutter pub get
flutter run --dart-define-from-file=env.json -d chromepip install requests google-auth
python scripts/ingest_pubmed_enhanced.py # ~50k articles, with Vertex AI embeddings
python scripts/ingest_pubmed_enhanced.py --no-embed # faster, keyword search onlyThe ingestion script is resumable — if interrupted, re-running skips already-indexed articles.
Why RAG instead of fine-tuning? Fine-tuning bakes knowledge into weights that go stale. RAG retrieves live knowledge from a corpus that can be updated, audited, and cited. For a medical system, auditability is not optional.
Why Qdrant? Qdrant is purpose-built for vector search with a clean REST API and a free cloud tier that handles this dataset comfortably. It replaced an earlier Elasticsearch implementation when the managed ES subscription ended.
Why force citations in the prompt? Without a hard constraint, LLMs blend retrieved facts with training-set knowledge in ways that are invisible to the user. The citation constraint makes the grounding verifiable — you can check every claim against its source.
Why 768 dimensions?
Vertex AI text-embedding-004 at 768 dims balances retrieval quality with storage cost. Qdrant's free tier holds this corpus at full precision.
- Answers are bounded by what is in the indexed corpus. If a topic is underrepresented in the dataset, retrieval quality suffers.
- Abstract-level search: full paper text is not indexed, only abstracts. Some nuance is lost.
- This is a research and education tool. It is not a medical device and does not provide medical advice.
MIT — fork, extend, deploy.