Chemical reaction extraction guide

Everything you need to run an extraction is in the workspace.

1. Add a source

   Upload a reaction image or PDF, or enter a public HTTPS image URL.

2. Choose a model connection

   Start with your free trial credits, or enter your own provider API key and vision model.

3. Extract and inspect

   Run the extraction, compare the result with the source, and correct structures before exporting.

Choose the matching source tab in the settings panel. The same workspace shows a preview before processing.

PDF processing can make several provider requests: one screening request per page, then extraction and validation for each matching page.

Use a vision-capable provider with an OpenAI-compatible chat-completions endpoint and JSON output. Enter these values in the left settings panel:

The workspace keeps the source and extraction results together. Review the Structured, Structures, and Raw JSON views, then use Copy JSON or Download to export.

Each experiment can include reactants, products, conditions, catalyst loading, yield and its measurement type, separate ee/er/dr values, notes, and confidence. Condition evidence links back to the original PDF page. Missing evidence is left empty or null. Check the model output against the document.

{
  "status": "completed",
  "results": [{
    "reactions": [{
      "reaction_id": "r1",
      "reactants": [{ "label": "1", "name": null, "smiles": "..." }],
      "products": [{ "label": "2", "name": null, "smiles": "..." }],
      "reagents": [], "solvents": [], "catalysts": [],
      "temperature": null, "time": null, "yield": null,
      "notes": [], "confidence": 0.8
    }]
  }],
  "meta": { "backend": "cloud-model-only", "validated": true }
}

The separate editor panel lets you sketch a molecule or load a SMILES string even before extraction. You can also open an extracted reactant or product in the Structures result view, correct it, and apply the updated SMILES to the result JSON.

The web app sends requests to its own Next.js route, which passes images and bounded PDF text evidence to the cloud-only Worker. For your own deployment, configure the web server with CHEMEAGLE_WORKER_URL and CHEMEAGLE_WORKER_ACCESS_TOKEN. Keep the Worker token on the server.

The Worker accepts POST /v1/extract with a JSON body like this:

{
  "provider": {
    "base_url": "https://your-provider.example/v1",
    "api_key": "YOUR_KEY",
    "api_key_header": "authorization",
    "model": "your-vision-model"
  },
  "source": {
    "mime_type": "image/png",
    "base64": "..."
  },
  "options": { "validate": true }
}

Direct Worker calls require x-chemeagle-access-token. The Worker accepts images; the web app renders and screens PDFs before sending image pages. See the Worker README for deployment details and the web README for local setup.

The original ChemEAGLE research implementation combines specialist extraction agents and tools across figures, tables, and text. The cloud workspace on this site is a lighter vision-model pipeline: it combines images with PDF text and runs a validation pass against the same evidence. Its results should not be treated as the original hybrid system's benchmark.