Skip to content

troubleshooting.md

Latest commit

 

History

History
160 lines (112 loc) · 3.73 KB

troubleshooting.md

File metadata and controls

160 lines (112 loc) · 3.73 KB
Raw

Troubleshooting

This page covers the failures most likely to occur during a Rowan autosearch run.

ROWAN_API_KEY not set

rowan-score requires a Rowan API key.

Fix:

cp .env.example .env

Then set ROWAN_API_KEY in .env. The scorer loads .env from the repository root.

Invalid SMILES

Symptom in iteration JSON:

{
  "score": null,
  "error": "RDKit failed to parse SMILES",
  "constraint_failures": ["invalid SMILES"]
}

Fix:

  • Check parentheses, ring closure numbers, charges, and aromatic atom casing.
  • Prefer canonical SMILES from RDKit or a trusted drawing tool.
  • Use rowan-score --dry-run for formatting checks before spending Rowan credits.

Metric Extraction Fails

Symptom:

{
  "score": null,
  "error": "Could not extract metric ..."
}

Fix:

  1. Open the latest runs/<run_id>/iterations/NNNN.json.
  2. Inspect candidates[*].object_data.
  3. Find the numeric value that matches the objective.
  4. Add its dot-path to runs/<run_id>/config.json as metric_path.
  5. Rescore the candidate or score a fresh baseline iteration.

Example:

{
  "metric_path": "solubilities.O.solubilities.0"
}

All Candidates Fail Constraints

Symptoms:

  • Candidate score values may be numeric.
  • satisfies_constraints is false.
  • constraint_failures lists the violated local descriptor limits.
  • rowan-state status does not treat them as best-so-far.

Fix:

  • Propose smaller or less lipophilic modifications.
  • Check whether the configured constraints reflect the project goal.
  • Consider easing constraints only if they were arbitrary or exploratory.
  • Do not chase a high score that is outside hard project constraints.

Rowan Submission Fails

Symptom:

{
  "score": null,
  "error": "Rowan submit failed: ..."
}

Fix:

  • Confirm ROWAN_API_KEY is valid.
  • Confirm workflow_type maps to an available Rowan SDK submitter.
  • Confirm required workflow_params are present.
  • Reduce --max-workers if failures look like rate limiting.
  • Retry only failed candidates when possible to avoid duplicate successful records.

Wrong Workflow Parameters

--workflow-param values are parsed as JSON when possible. Lists and objects must be quoted for the shell:

--workflow-param solvents='["water"]'
--workflow-param temperatures='[298.15]'
--workflow-param pocket='[[0,0,0],[12,12,12]]'

Unquoted values are usually parsed as strings or numbers:

--workflow-param method=kingfisher
--workflow-param exhaustiveness=8

If Rowan rejects a parameter, compare workflow_params in config.json with the current Rowan SDK submitter signature.

Report Does Not Show A Winner

The report uses constraint-aware best-so-far logic. It will not show a winner if no candidate has both:

  • a numeric score
  • satisfies_constraints: true

Fix scoring errors or constraints first, then rerun:

uv run rowan-report --run <run_id>

rowan-propose Output Is Chemically Odd

rowan-propose is a brainstorming helper. It uses simple RDKit transformations and may emit candidates that are syntactically valid but strategically weak.

Fix:

  • Treat output as suggestions, not an auto-submit list.
  • Filter for chemistry that matches the objective.
  • Check constraints locally through rowan-score --dry-run or by estimating descriptor changes before scoring.

Duplicate Or Bad Iteration Files

The run history is intended to be append-only. Avoid rewriting old iterations.

Acceptable cleanup:

  • Remove an accidental duplicate candidate before further analysis.
  • Fix a mistaken decision value.
  • Drop failed duplicates created by a retry, if the successful result already exists elsewhere.

When in doubt, append a new iteration and explain the correction in the --rationale.