SubsettingAgent

AI agent that helps genebank users build subsets of accessions (seed samples) combining passport data, traits, climate indicators and research papers. Accessions come from an uploaded Excel/CSV file (local mode) or from Genesys PGR (Genesys mode).

Architecture

app.py                  Gradio chat interface (multimodal: text + files)
agent/
  subsetting_agent.py   LLM tool-calling loop (LiteLLM) that dispatches skills
  prompts.py            System prompt template
  skill_registry.py     Auto-discovery of skills under skills/
skills/
  base.py               Skill contract (name, description, parameters, run)
  <skill_name>/         One folder per skill: SKILL.md + skill.py
  list_accessions/      Load the first list (local Excel/CSV, or Genesys PGR with structured criteria)
  passport_filter/      Filter / describe / reset the Candidate list by passport columns
  climate_analysis/     list_indicators / filter / cluster by climate indicators (Subsetting API)
  traits_analysis/      detect / fetch (Genesys datasets) / group by trait data -> cluster_traits
  export_list/          Export Candidate or Original list as CSV (attached to the chat)
  arguments.py          Tolerant parsing of tool arguments (lists, ints, bools, JSON text)
core/
  config.py             Settings loaded from .env (pydantic-settings)
  session.py            tmp/<session>/inputs|outputs folders
  state.py              SessionState: Original list, Candidate list, mode, log
  formatter.py          Activity summary + 15-row preview + CSV exports
  logger.py             Logging setup
sdks/
  genesys/              Genesys PGR client (accession listing, crops, datasets/descriptors/observations) - see sdks/genesys/README.md
  subsetting/           Subsetting API client (indicators, indicator data, clusters, cellid) - see sdks/subsetting/README.md
scripts/
  genesys_smoke.py      Manual check of the Genesys SDK against the real API
  genesys_traits_smoke.py  Manual check of the Genesys trait/dataset endpoints (prints raw rows)
  subsetting_smoke.py   Manual check of the Subsetting SDK against the real API
tests/                  pytest suite

Requirements

• Python 3.10
• uv
• An LLM reachable through LiteLLM (Ollama with llama3.1:8b by default)

Setup

uv sync                      # creates .venv and installs dependencies
cp .env.example .env         # then edit the values you need

Run

uv run app.py

Open http://localhost:7860 (or the host/port set in .env).

Test

uv run --group dev pytest

Business rules in short

• Original list: first list of accessions loaded in the session; never changes.
• Candidate list: working list returned to the user; every skill narrows or annotates it (not cumulative). It carries the columns criteria_passport, criteria_traits, criteria_research, criteria_climate, cluster_traits, cluster_research and cluster_climate.
• Every answer contains a summary of the activities performed and a preview of the first 15 rows of the Candidate list. Full CSV exports are generated on request.