Local-First AI Research & Bioinformatics Hub
To empower autonomous AI coding and scientific agents without token leakage, rate limits, or slow external web scraping, Elijah built a dedicated, private research and bioinformatics search engine inside Proxmox CT 107. By uniting DuckDB inverted BM25 full-text indexing, Google-style Zoekt trigram code indexing, local NCBI BLAST+ nucleotide/protein alignment, and offline Kiwix ZIM archives behind the Model Context Protocol (FastMCP), local AI models can execute thousands of deep literature, repository, and genomics lookups in single-digit milliseconds.
Indexed Corpus & Engine Status
Real-time status proxied directly from CT 107:8080 via the Proxmox homelab API backend. When operating offline or outside the local network, verified baseline snapshots are displayed.
Agent-to-Engine Retrieval Topology
How agent reasoning loops interact with CT 107 over standardized Model Context Protocol tools.
The 4 Pillars of the Research Hub
Eliminating external API dependence through purpose-built local database and indexing engines.
DuckDB Full-Text BM25 Literature Engine
Public paper search APIs are notoriously slow, rate-limited, and lack rich full-text search syntax. We ingested over 71,500 arXiv preprints (spanning machine learning, artificial intelligence, quantitative biology, and computational neuroscience) directly into a columnar DuckDB database.
- Inverted Index: Pre-built BM25 scoring across titles, abstracts, and author affiliations.
- Sub-Millisecond Response: Average query latency of 2.8 ms, allowing agents to test 20+ query reformulations in seconds.
- Zero Per-Query Cost: Zero external LLM token overhead or paid search engine credits.
Zoekt Sub-10ms Trigram Code Search
When an AI agent investigates an algorithm, reading papers isn't enough—it needs ground-truth source code. We deployed Zoekt (Google's open-source trigram search engine) indexed across 59 upstream scientific and ML repositories.
- 65 Pre-Indexed Shards: PyTorch, Hugging Face Transformers, Biopython, AlphaFold tools, RDKit, and local project repos.
- Instant Regex & Symbols: Exact symbol definitions, function signatures, and imports returned in under 8 ms.
- No Repository Cloning: Agents query millions of lines of upstream source code without polluting local disk or agent context windows.
Local NCBI BLAST+ & RefSeq Genomics
Web-based NCBI BLAST queues can take minutes during peak hours and leak proprietary sequence fragments over public networks. CT 107 houses a complete, locally managed NCBI BLAST+ 2.15+ toolchain with optimized binary indices.
- 19,625 Viral Genomes: RefSeq and GenBank viral nucleotide/protein reference databases formatted with
makeblastdb. - Relational Taxdump: Integrated NCBI Taxonomy relational SQLite database linking Accession IDs to lineage trees in real time.
- Deterministic Alignment: E-values, bitscores, and query coverages returned in 40–120 ms directly to AI reasoning agents.
FastMCP Agent Protocol Layer
To make all three engines natively accessible to modern AI assistants, we unified them behind the Model Context Protocol (FastMCP) running over high-speed HTTP SSE and JSON-RPC.
- Agent Tool Interface: Clean function calling definitions (
search_arxiv,search_code,run_blast,query_archive). - Single Reasoning Loop: An agent can search literature, verify the algorithm's code implementation, and align biological data within a single continuous thought stream.
- Interoperable: Natively compatible with Google Antigravity, Claude Code, Cursor, and custom LangGraph/CrewAI agents.
How an AI Agent Uses CT 107 in Real Time
A concrete trace of an autonomous research session executing across local tools.