Project: Bioinfo Bot
Backed by two knowledge sources: a curated corpus of ~15,000 BioRxiv bioinformatics preprints (RAG), and an in-house PubMed MCP server for live peer-reviewed search. Ask it anything...
e.g. "explain genome assembly like I'm 5" or "find recent papers on CRISPR base editing"
Powered by Claude Sonnet 4.6 · Voyage AI embeddings · pgvector · FastMCP · Google Cloud
What
Bioinfo Agent is a proof of concept agentic research assistant for bioinformatics. It draws on two complementary knowledge sources: a curated BioRxiv preprint corpus for depth on niche computational biology topics, and a live PubMed MCP server for peer-reviewed literature currency. Claude Sonnet 4.6 acts as the agent, deciding which sources to consult — and in what combination — based on the question.
Question → Answer flow
- Your question is embedded via Voyage AI (
voyage-4-large, 1024 dimensions) - A cosine similarity search runs against ~15,000 BioRxiv abstracts stored in Cloud SQL (pgvector) — only papers within a relevance threshold are kept
- Claude receives your question + any relevant BioRxiv context + prior PubMed results (if this is a follow-up), then decides whether to call the PubMed MCP server
- If Claude calls PubMed, the request is authenticated via GCP service-to-service identity tokens and routed to a FastMCP server on Cloud Run, which queries NCBI's E-utilities API
- Claude synthesises a response citing papers inline (Author et al., year), streamed back as Server-Sent Events (SSE) and rendered word-by-word in the browser
GCP architecture
| Service | Role |
|---|---|
| Cloud Run (Django) | Main web app — handles requests, RAG retrieval, Claude streaming |
| Cloud Run (FastMCP) | In-house PubMed MCP server — wraps NCBI E-utilities,
exposes search_pubmed, get_paper_metadata,
find_related_papers
|
| Cloud SQL (PostgreSQL) | Stores BioRxiv paper embeddings via the pgvector extension |
| Secret Manager | Stores API keys (Anthropic, Voyage AI) |
| Artifact Registry + Cloud Build | Container image storage and CI/CD pipeline |
v1 → v2 → v3
| Agent v1 (2023) | Agent v2 (2025) | Agent v3 (2026) | |
|---|---|---|---|
| LLM | Claude Sonnet 2.x | Claude Sonnet 4.6 | Claude Sonnet 4.6 |
| Embeddings | OpenAI ada-002 | Voyage AI voyage-3-large | Voyage AI voyage-4-large |
| Vector DB | Pinecone (managed) | pgvector on Cloud SQL | pgvector on Cloud SQL |
| Chat state | Global Python list (shared) | Django sessions (per-user cookie) | Per-page-load UUID (in-memory, tab-isolated) |
| Knowledge sources | BioRxiv RAG only | BioRxiv RAG only | BioRxiv RAG + live PubMed MCP (agent decides) |
| PubMed tools | — | — | search, get metadata by PMID, find related papers |
| MCP auth | — | — | GCP service-to-service identity tokens |
| References | None | Top 5 BioRxiv papers per response | BioRxiv preprints + PubMed live results, badged by source |
Why
Demonstrate that an AI agent backed by structured retrieval and live tool use can meaningfully compress the time cost of literature review. The BioRxiv corpus alone represents more than 15 years of daily reading at 3 papers/day — PubMed extends that to tens of millions of peer-reviewed records, searched on demand.