Open Source Research Tool

Agentic evidence-grounded research assistant for biology

Alvessa integrates entity recognition, validated tool orchestration, and statement-level verification to deliver trustworthy, provenance-preserving answers to biomedical questions.

Built for trustworthy biomedical reasoning

Entity Recognition

Purpose-built ensemble combining LLMs, Flair, GLiNER, and regex patterns to identify genes, variants, drugs, and miRNAs.

Validated Tool Catalog

Pre-defined, tested tools for 20+ databases including ChEMBL, UniProt, Reactome, GWAS Catalog, and AlphaFold-ensuring consistent, reproducible results.

Statement-Level Verification

Every claim is evaluated against retrieved evidence and classified as supported, partially supported, or unsupported-with explanations surfaced to users.

Evidence Provenance

Direct citations link every statement to source records. Explore underlying data and verify claims independently.

Interactive Outputs

Rich HTML reports with 3D protein viewers, molecule renderings, and expandable evidence panels for deep exploration.

GenomeArena Benchmark

720 curated questions spanning 8 categories to evaluate genomic reasoning capabilities-freely available for community use.

Multi-agent pipeline with verification loop

User Query
Entity Recognition
Orchestration Loop
Tool Selection
Evidence Collection
Verification Loop
Answer Generation
Verification Agent
Verified Output
Feedback loops enable refinement

Example Outputs

Resources & Documentation

Reference

@article{sokolova2025alvessa, title = {An Evidence-Grounded Research Assistant for Functional Genomics and Drug Target Assessment}, author = {Sokolova, Ksenia and Kosenkov, Dmitri and Nallamotu, Keerthana and Vedula, Sanketh and Sokolov, Daniil and Sapiro, Guillermo and Troyanskaya, Olga G}, journal = {bioRxiv}, year = {2025}, month = dec, doi = {10.64898/2025.12.30.697073}, url = {https://www.biorxiv.org/content/10.64898/2025.12.30.697073v1}, note = {Preprint} }