v0.2.0

What's New

The gains came from better data and from fixing bugs, not from cleverer scoring.

• Vendor catalogues. reagent build-catalogue turns an Enamine or eMolecules download into usable stock. The biggest win here: ZINC alone leaves 3 of 10 hard targets unsolved, ZINC plus eMolecules solves all 10, and the routes get longer rather than shorter.
• Runs on 8 GB of RAM. --hashed-stock cuts peak memory from 4.91 GB to 0.63 GB with identical results, so searches no longer get killed.
• Three more search algorithms. Retro*, DFPN and breadth-first, via --algorithm.
• A seventh objective, construction, which scores how much of the molecule a route builds instead of buys.
• Faster evaluation. --jobs plans several targets at once, sized to your free memory.

Results

	v0.1.0	v0.2.0
Solve-rate [Easy set]	0.90	1.00
Solve-rate [Hard set]	not run	1.00
Peak memory	4.91 GiB	0.63 GiB
Tests	37	127

Fixed

• Objectives are now compared on a level footing, so weights express what you prefer rather than which objective happens to have the widest range.
• Safety no longer penalises a route just for having more steps.
• The same inputs now give the same answer. Ties were being decided by the order routes arrived in, which also cost real quality. Fixing it cut routes that buy a near-finished molecule from 5 to 3.

What was tested

Four ideas were built, measured, and did not help:

• steering the search by hazard
• picking the most balanced route instead of the best-weighted one
• weighting buy-versus-build, and
• combining expansion policies.

The README records what each was worth.

Getting Started

mise install && mise run install
download_public_data data
reagent build-stock-cache
reagent plan "CC(=O)Oc1ccccc1C(=O)O"

Stock files are not attached, however the README has the exact commands to build them. reagent-0.2.0-measurements.tar.gz holds the raw output behind every number above.

Full Changelog: v0.1.0...v0.2.0

ReAgent v0.1.0

First tagged release.

• Retrosynthetic route planning via AiZynthFinder (single-step model + MCTS).
• Deterministic multi-objective feature layer: feasibility, availability, cost, safety, sustainability, efficiency.
• Specialist LLM evaluator agents (Anthropic + local Ollama), hybrid score-plus-explain mode, confidence gate.
• Real safety data via PubChem GHS (--ghs); Brenk structural-alert screen offline.
• RAG precedent retrieval over USPTO reaction fingerprints.
• Episodic memory and feedback-driven objective-weight tuning.
• Evaluation and agent-accuracy harnesses. 37 tests, ruff clean.