LeMaterial

Quickstart

This page gets you from zero to your first extracted synthesis in under 10 minutes.

Option 1 — Interactive notebook (recommended for beginners)

Open the tutorial notebooks in Jupyter:

uv run jupyter lab examples/notebooks/tutorials/

Start with Tutorial 3 — Batch extraction with the CLI, which gets you results in one command, or Tutorial 1 if you would rather read the published dataset than run anything. Tutorial 4 walks a fixed example paper through the pipeline a stage at a time when you want to see the internals. Each step prints its result before the next one starts, so you can see exactly what the models returned and stop wherever you like.

The Tutorials index lists all seven, with the API keys and rough cost each one needs.

Option 2 — Command line (one paper)

# The lemat-synth CLI reads .env itself — no need to source anything.
# Just make sure .env at the repo root contains GEMINI_API_KEY=...

# Extract synthesis from one text file
lemat-synth extract my_paper.txt

# Extract from a PDF (text is extracted automatically with Docling)
lemat-synth extract my_paper.pdf

# Results are saved to results/<paper-name>/<material>.json

Option 3 — Command line (batch)

lemat-synth batch /path/to/my_papers/ output_dir=results/

Settings are given as key=value pairs after the folder — there are no --flags. Add max_papers=5 to process only the first 5 papers as a test:

lemat-synth batch /path/to/my_papers/ output_dir=results/ max_papers=5
Tip
Always do a max_papers=5 run first. Cost and runtime scale with the number of materials, not papers, so a folder of catalysis papers can be several times more expensive than the same number of single-material papers.

Option 4 — Python API (one paper, no config files)

import json
import os
from dotenv import load_dotenv

load_dotenv()

from llm_synthesis.utils.dspy_utils import get_llm_from_name
from llm_synthesis.utils import configure_dspy
from llm_synthesis.transformers.material_extraction.dspy_extraction import (
    DspyTextExtractor, make_dspy_text_extractor_signature,
)
from llm_synthesis.transformers.synthesis_extraction.dspy_synthesis_extraction import (
    DspySynthesisExtractor, make_dspy_synthesis_extractor_signature,
)

configure_dspy("gemini-2.0-flash")
paper_text = open("my_paper.txt").read()

# Step 1: find materials
material_extractor = DspyTextExtractor(
    signature=make_dspy_text_extractor_signature(
        instructions="Extract all synthesized materials as chemical formulas.",
        output_description="Comma-separated list of material formulas.",
    ),
    lm=get_llm_from_name("gemini-2.0-flash", model_kwargs={"temperature": 0.0}),
)
materials = [
    m.strip()
    for m in material_extractor.forward(input=paper_text).split(",")
    if m.strip()
]
print("Materials:", materials)

# Step 2: extract synthesis for the first material
synthesis_extractor = DspySynthesisExtractor(
    signature=make_dspy_synthesis_extractor_signature(
        instructions="Extract the complete synthesis procedure."
    ),
    lm=get_llm_from_name("gemini-2.0-flash", model_kwargs={"temperature": 0.0}),
)
synthesis = synthesis_extractor.forward(input=(paper_text, materials[0]))
print(json.dumps(synthesis.model_dump(), indent=2))

Understanding the output

Each result file looks like this (simplified):

{
  "material": "Fe2O3",
  "synthesis": {
    "target_compound": "Fe2O3",
    "synthesis_method": "hydrothermal",
    "starting_materials": [
      {"name": "FeCl3", "amount": 1.62, "unit": "g", "purity": "98%"}
    ],
    "steps": [
      {"step_number": 1, "action": "dissolve", "conditions": {"temperature": 25, "temp_unit": "C"}},
      {"step_number": 2, "action": "heat",    "conditions": {"temperature": 180, "temp_unit": "C", "duration": 12, "time_unit": "h"}}
    ]
  },
  "evaluation": {
    "scores": {
      "structural_completeness_score": 4.5,
      "material_extraction_score": 4.0,
      "overall_score": 4.2,
      "overall_reasoning": "Faithful to the source; drying time not stated in the paper."
    },
    "confidence_level": "high",
    "missing_information": [],
    "extraction_errors": []
  }
}

For a full explanation of every field, see Output Format.

Next steps

GoalResource
Process many paperslemat-synth batch or deployment scripts
Also extract performance plotslemat-synth batch ... with_performance=true
Use the Python API for a custom workflowPython API guide
Change the LLMConfiguration guide
See every CLI settingCLI Reference
Run dataset-scale or multi-LLM jobsConfiguration guide
Prev
Installation
Next
Dataset Access