Remove annotation hints from holistic extractor prompt

AGENTS.md

@@ -0,0 +1,23 @@
+## Developer Workflow
+Follow SOLID principles.
+Always ensure that tests pass before committing.
+
+
+## Coding Standards
+- Keep routers thin; logic in `services/`.
+- Repositories are the only DB touchpoint from services.
+- DTOs (schemas) are versioned; never leak ORM models.
+- Return 201 on creates, 202 on async accepted; 429 on limits; 422 on validation.
+- Docstrings: Google style; type hints everywhere.
+- Avoid defining `__all__` unless absolutely necessary, and never use `from module import *`; prefer explicit imports (ruff F403/F405 guard against this).
+- Keep package `__init__.py` files empty aside from optional module docstrings; avoid re-exporting symbols from them.
+- Never add `from __future__ import annotations`; native Python 3.12 typing is required.
+- Prefer native union syntax (e.g., `str | None`) instead of `typing.Optional[...]` or `typing.Union`.
+
+## Development Workflow
+
+- All PRs require: ✅ lint, ✅ type-check, ✅ tests.
+- Before committing, run all relevant project checks locally and ensure they pass.
+- Write concise PR descriptions (why + what), not just code diffs.
+- Instructions from user may be asked in any language, but code and response should be in English.
+- Do not use `docker compose up -d` (or any detached Compose command) in development-facing scripts or docs.

README.md

@@ -143,18 +143,21 @@ print clear instructions for manual download.
        --dataset-dir data/images --out results
    ```
 
-   This will create a JSONL file in `results/` whose name begins with the
-   model name and includes the current timestamp.  Each line contains the
-   image file name, the extracted text and any additional metadata returned by
-   the model.  The extractor includes a template prompt that instructs the
-   model to convert the chemical diagram into a SMILES string.  You may
+   This command now performs **holistic page analysis**.  Each JSONL line
+   corresponds to a single patent page and stores the prompt, raw model
+   response and any parsed JSON payload containing `structures` (with
+   model-assigned identifiers and SMILES strings) and `reactions`
+   (reactant/product identifier sets plus optional conditions).  The default
+   prompt keeps the model blind to MolMole annotations so it must discover the
+   number of structures and reactions directly from the image.  You may
    customise the prompt or provide an alternative generation function by
    editing `extractor.py`.
 
 5. **Evaluate predictions**: use the evaluator script to compare your
-   predictions against the ground truth SMILES strings.  The evaluator
-   canonicalises SMILES using RDKit (if available) and computes accuracy
-   using both SMILES and InChI‑Key matching:
+   predictions against the ground truth SMILES strings and reaction graphs.
+   The evaluator canonicalises SMILES using RDKit (if available), matches
+   predictions per structure identifier, and reports SMILES/InChI accuracy,
+   mean Tanimoto similarity and reaction precision/recall/F1:
 
    ```bash
    python -m molmole_research.evaluator --pred results/gpt-4o-vision_*.jsonl \

src/molmole_research/__init__.py

@@ -1,23 +1 @@
-"""Top‑level package for MolMole OCSR research environment.
-
-This package contains modules to download the MolMole benchmark dataset,
-extract chemical structures from images using vision‑language models,
-evaluate predictions against ground truth and orchestrate experiments.
-
-The package is designed to be executed via the command line, for example:
-
-    python -m molmole_research.downloader --help
-    python -m molmole_research.extractor  --help
-    python -m molmole_research.evaluator  --help
-    python -m molmole_research.runner     --help
-
-"""
-
-__all__ = [
-    "downloader",
-    "extractor",
-    "evaluator",
-    "runner",
-]
-
-__version__ = "0.1.0"
+"""Top-level package for MolMole OCSR research environment."""