Build a Local Chemistry Assistant with Ollama + RDKit

1. Prepare Python and Ollama

All platforms

2. Install isolated Python dependencies

macOS

python3.11 -m venv .venv
.venv/bin/python -m pip install rdkit==2025.3.1
.venv/bin/python -m pip freeze

3. Install isolated Python dependencies

Linux

python3.11 -m venv .venv
.venv/bin/python -m pip install rdkit==2025.3.1
.venv/bin/python -m pip freeze

4. Install Python dependencies on Windows

Windows

py -3.11 -m venv .venv
.\.venv\Scripts\python.exe -m pip install rdkit==2025.3.1
.\.venv\Scripts\python.exe -m pip freeze

5. Download the local explanation model

All platforms

ollama --version
ollama pull qwen2.5:3b
ollama list

6. Save the Python bridge

All platforms

"""RDKit calculations followed by explanation using a local Ollama model.

Sources: https://www.rdkit.org/docs/GettingStartedInPython.html
         https://docs.ollama.com/api/chat
This example has been source-reviewed, but has not been executed.
"""
import argparse
import json
import urllib.error
import urllib.request

from rdkit import Chem, rdBase
from rdkit.Chem import Descriptors, rdMolDescriptors


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("smiles")
    parser.add_argument("--model", default="qwen2.5:3b")
    parser.add_argument("--prompt", default="Explain the supplied molecular descriptors and their limits.")
    args = parser.parse_args()
    if not args.smiles.strip():
        parser.error("Empty SMILES; no model request was sent.")
    molecule = Chem.MolFromSmiles(args.smiles)
    if molecule is None:
        parser.error("Invalid SMILES; no model request was sent.")
    results = {
        "smiles": args.smiles,
        "rdkit_version": rdBase.rdkitVersion,
        "formula": rdMolDescriptors.CalcMolFormula(molecule),
        "molecular_weight_g_mol": Descriptors.MolWt(molecule),
        "tpsa_square_angstrom": rdMolDescriptors.CalcTPSA(molecule),
    }
    print(json.dumps(results, ensure_ascii=False, indent=2), flush=True)
    request = urllib.request.Request(
        "http://127.0.0.1:11434/api/chat",
        data=json.dumps({
            "model": args.model,
            "stream": False,
            "messages": [
                {"role": "system", "content": "Explain only the supplied RDKit results. Preserve their numbers and units. Do not invent measurements, toxicity, efficacy or additional calculations. Reply in the user's language."},
                {"role": "user", "content": args.prompt + "\nRDKit results:\n" + json.dumps(results)},
            ],
        }).encode("utf-8"),
        headers={"Content-Type": "application/json"},
        method="POST",
    )
    try:
        opener = urllib.request.build_opener(urllib.request.ProxyHandler({}))
        with opener.open(request, timeout=180) as response:
            body = json.load(response)
        message = body.get("message", {}).get("content")
        if not isinstance(message, str) or not message.strip():
            raise ValueError("Ollama did not return message.content.")
        print(message)
    except (urllib.error.URLError, ValueError, TimeoutError) as error:
        parser.exit(1, "Local Ollama request failed; RDKit results remain available above. " + str(error) + "\n")


if __name__ == "__main__":
    main()

7. Run the aspirin workflow

All platforms

.venv/bin/python local_chemistry_assistant.py "CC(=O)Oc1ccccc1C(=O)O" --prompt "Explain the aspirin results computed by RDKit. Preserve formula, mass and TPSA with units; distinguish computed descriptors from experimental observations."

8. Reject malformed input

All platforms

.venv/bin/python local_chemistry_assistant.py "not-a-smiles"