CLI Reference

The lemat-synth command-line tool lets you extract structured synthesis procedures from materials science papers without writing any Python code.

lemat-synth extract <paper>   [key=value ...] # Extract from a single paper
lemat-synth batch   <folder>  [key=value ...] # Extract from a folder of papers

Both commands accept the same key=value overrides, which can be used to change models, prompts, output paths, and other settings. See Quick Reference: All Arguments below.

By default, extraction runs a single LLM pass over the whole paper. Pass domain=catalysis (or superconductors, electrochemistry) to filter figures down to domain-relevant plots, add with_performance=true to also extract and link plot data to materials (requires ANTHROPIC_API_KEY), and override any prompts.* key, e.g., prompts.synthesis_instructions="...", to customize the extraction instructions.

Quick Reference: All Arguments

All settings — models, prompts, domain, output path — live in examples/config/cli.yaml. You can override any of them directly on the command line using Hydra key=value syntax.

Argument	Type	Default	Description
Models & API
`synthesis_model`	string	`gemini/gemini-3.5-flash-lite`	Main extraction model (LiteLLM format)
`material_model`	string	`gemini/gemini-3.5-flash-lite`	Fast model for material-list extraction
`judge_model`	string	(mirrors `synthesis_model`)	Quality evaluation model
`linker_model`	string	`gemini/gemini-3.1-pro-preview`	Links plots to materials (requires `with_performance=true`)
`plot_model`	string	`claude-sonnet-4.6`	Claude model for plot data extraction (requires `with_performance=true`)
`api_base`	string	`null`	Custom API base URL, e.g. `https://openrouter.ai/api/v1`
`synthesis_api_key_env`	string	`null`	Env var name holding synthesis model API key
`material_api_key_env`	string	`null`	Env var name holding material model API key
`judge_api_key_env`	string	(mirrors `synthesis_api_key_env`)	Env var name for judge model
`linker_api_key_env`	string	`null`	Env var name for linker model
Pipeline Behavior
`domain`	choice	`generic`	Plot filtering: `generic`, `catalysis`, `superconductors`, `electrochemistry`
`with_performance`	bool	`false`	Extract performance data and link to materials (requires Claude API key)
`output_dir`	path	`results`	Output directory for results
`pdf_extractor`	choice	`docling`	PDF extraction backend: `docling` (local) or `mistral` (API-based)
`figure_segmenter`	choice	`dino`	Figure segmentation: `dino` or `florence`
`florence_repo_id`	string	`amayuelas/plot-visualization-florence-2-lora-32`	HuggingFace LoRA adapter ID (when `figure_segmenter=florence`)
Batch Only
`max_papers`	int	`null`	Maximum papers to process (`null` = all)
`skip_existing`	bool	`true`	Skip papers already in output directory
`max_papers_parallel`	int	`4`	Concurrent papers to process
Prompts
`prompts.synthesis_system`	string	(see below)	System message for synthesis extraction
`prompts.synthesis_instructions`	string	(see below)	Task instructions for synthesis extractor
`prompts.material_instructions`	string	(see below)	Task instructions for material extractor
Other prompt keys	string	(see below)	See Customising prompts

Examples

Basic usage

# Uses all defaults from examples/config/cli.yaml
lemat-synth extract paper.txt

# Custom output folder
lemat-synth extract paper.txt output_dir=my_results/

Common customizations

# Use a different synthesis model
lemat-synth extract paper.txt synthesis_model=anthropic/claude-sonnet-4-6

# Domain-specific plot filtering (catalysis, superconductors, or electrochemistry)
lemat-synth extract paper.txt domain=catalysis

# Extract performance data and link plots to materials (requires ANTHROPIC_API_KEY)
lemat-synth extract paper.txt with_performance=true

# Use Mistral OCR for better PDF extraction (requires MISTRAL_API_KEY)
lemat-synth extract paper.pdf pdf_extractor=mistral

# Override the synthesis extraction prompt (inner 'single quotes' are required
# here because the value contains a comma — see "Customizing Prompts" below)
lemat-synth extract paper.txt \
    "prompts.synthesis_instructions='Extract only the primary synthesis route, ignoring alternative procedures.'"

Advanced: OpenRouter with multiple API keys

# All models through OpenRouter (Gemini Flash for synthesis, Claude for performance)
lemat-synth extract data/cipollone_2022.pdf \
    api_base="https://openrouter.ai/api/v1" \
    pdf_extractor=mistral \
    material_model="openrouter/google/gemini-3.1-pro-preview" \
    material_api_key_env=GEMINI_API_KEY \
    synthesis_model="openrouter/google/gemini-3-flash-preview" \
    synthesis_api_key_env=GEMINI_API_KEY \
    linker_model="openrouter/google/gemini-3.1-pro-preview" \
    linker_api_key_env=GEMINI_API_KEY \
    output_dir="results/"

Caution

The _api_key_env arguments must not contain the actual API key: they must be the name of an environment variable that holds the key. For example, material_api_key_env=GEMINI_API_KEY means that the material model set to Gemini will use the API key stored in the environment variable GEMINI_API_KEY. See API key environment variables for more details.

Advanced: Extract with performance linking (OpenRouter)

Extract synthesis procedures and link extracted plot data to synthesized materials:

# Same as above, plus performance extraction using Claude via OpenRouter
lemat-synth extract data/cipollone_2022.pdf \
    api_base="https://openrouter.ai/api/v1" \
    material_model="openrouter/google/gemini-3.1-pro-preview" \
    material_api_key_env=GEMINI_API_KEY \
    synthesis_model="openrouter/google/gemini-3-flash-preview" \
    synthesis_api_key_env=GEMINI_API_KEY \
    linker_model="openrouter/google/gemini-3.1-pro-preview" \
    linker_api_key_env=GEMINI_API_KEY \
    plot_model="openrouter/anthropic/claude-sonnet-4.6" \
    output_dir="results/" \
    with_performance=true

Caution

The _api_key_env arguments must not contain the actual API key: they must be the name of an environment variable that holds the key. For example, material_api_key_env=GEMINI_API_KEY means that the material model set to Gemini will use the API key stored in the environment variable GEMINI_API_KEY. See API key environment variables for more details.

Batch Processing

# Basic — processes all papers in folder
lemat-synth batch papers/

# Quick test run (first 5 papers only)
lemat-synth batch papers/ max_papers=5

# Custom output folder and domain filtering
lemat-synth batch papers/ \
    output_dir=results/catalysis/ \
    domain=catalysis

# Re-process everything (skip_existing=false)
lemat-synth batch papers/ skip_existing=false

# Reduce parallelism to avoid rate limits
lemat-synth batch papers/ max_papers_parallel=2

# Use Mistral OCR for all PDFs
lemat-synth batch papers/ pdf_extractor=mistral

# Powerful models, catalysis domain, resume if interrupted
lemat-synth batch papers/ \
    synthesis_model=gemini/gemini-2.5-pro \
    material_model=anthropic/claude-opus-4-7 \
    domain=catalysis \
    skip_existing=true \
    max_papers_parallel=2 \
    output_dir="results/catalysis/"

# Different models through OpenRouter
lemat-synth batch papers/ \
    synthesis_model=openrouter/google/gemini-3-flash-preview \
    material_model=openrouter/google/gemini-3.1-pro-preview \
    judge_model=openrouter/anthropic/claude-sonnet-4.6 \
    api_base=https://openrouter.ai/api/v1

Configuration Details

All arguments in the Quick Reference table above can be overridden from the command line. Defaults are read from examples/config/cli.yaml.

Model strings

Model strings follow the LiteLLM convention: {provider}/{model-name}. Common providers and models:

gemini/gemini-3.5-flash-lite                      # Google Gemini
gemini/gemini-3.1-pro-preview
gemini/gemini-2.5-pro
anthropic/claude-sonnet-4-6                      # Anthropic Claude
anthropic/claude-opus-4-7
openai/gpt-4o                                    # OpenAI
openai/gpt-4o-mini
mistral/mistral-large                            # Mistral
openrouter/google/gemini-3-flash-preview          # OpenRouter (requires api_base + key)
openrouter/google/gemini-3.1-pro-preview
openrouter/anthropic/claude-sonnet-4.6
openrouter/deepseek/deepseek-v3.2
openrouter/qwen/qwen3.5-35b-a3b
openrouter/moonshotai/kimi-k2.5

When using OpenRouter models, always set api_base=https://openrouter.ai/api/v1.

API key environment variables

By default LiteLLM auto-detects API keys from standard environment variables:

gemini/* → GEMINI_API_KEY
anthropic/* → ANTHROPIC_API_KEY
openai/* → OPENAI_API_KEY
etc.

Use the *_api_key_env arguments to override this — useful for OpenRouter key slots or when multiple keys exist for the same provider.

# Example: different OpenRouter keys for different models
lemat-synth batch papers/ \
    synthesis_model=openrouter/qwen/qwen3.5-35b-a3b \
    synthesis_api_key_env=OPENROUTER_QWEN_API_KEY \
    linker_model=openrouter/moonshotai/kimi-k2.5 \
    linker_api_key_env=OPENROUTER_KIMI_API_KEY \
    api_base=https://openrouter.ai/api/v1

Caution

The _api_key_env arguments must not contain the actual API key: they must be the name of an environment variable that holds the key. For example, material_api_key_env=OPENROUTER_QWEN_API_KEY means that the material model set to QWEN will use the API key stored in the environment variable OPENROUTER_QWEN_API_KEY.

Allowed Environment Variables

Add these to your .env file (automatically loaded at runtime):

Variable	When required	Example use
`GEMINI_API_KEY`	Using `gemini/*` models	Default synthesis/material models
`ANTHROPIC_API_KEY`	Using Claude models or `with_performance=true`	`synthesis_model=anthropic/claude-sonnet-4-6`
`OPENAI_API_KEY`	Using `openai/gpt-*` models	`plot_model=openai/gpt-4o`
`MISTRAL_API_KEY`	Using Mistral models or `pdf_extractor=mistral`	`pdf_extractor=mistral` for better OCR
`OPENROUTER_QWEN_API_KEY`	Using Qwen via OpenRouter	`synthesis_model=openrouter/qwen/qwen3.5-35b-a3b`
`OPENROUTER_KIMI_API_KEY`	Using Kimi via OpenRouter	`linker_model=openrouter/moonshotai/kimi-k2.5`
`OPENROUTER_DEEPSEEK_API_KEY`	Using DeepSeek via OpenRouter	`synthesis_model=openrouter/deepseek/deepseek-v3.2`

PDF and figure processing

Argument	Options	When to use
`pdf_extractor`	`docling` (default)	Local, no API key required
	`mistral`	Better for scanned/low-quality PDFs (requires MISTRAL_API_KEY)
`figure_segmenter`	`dino` (default)	Fast, 28-class detection
	`florence`	More accurate, binary quantitative/qualitative classification
`florence_repo_id`	HuggingFace repo ID	LoRA adapter for Florence (only when `figure_segmenter=florence`)

Domain filtering (when `with_performance=true`)

Domain	Figures kept	Use case
`generic`	All figures (no filtering)	Default for multi-domain papers
`catalysis`	Conversion/selectivity vs temperature curves	Catalysis materials
`superconductors`	Resistivity ρ(T) and resistance R(T) plots	Superconductor data
`electrochemistry`	Current/capacitance vs voltage curves	Battery/electrochemistry materials

Batch processing options

Argument	Default	Purpose
`max_papers`	`null`	Stop after N papers (useful for test runs)
`skip_existing`	`true`	Resume from last run; set to `false` to reprocess all
`max_papers_parallel`	`4`	Concurrent papers; lower if hitting rate limits

Customizing Prompts

Every prompt used during extraction can be customized. The full set of prompt keys is in examples/config/cli.yaml under the prompts: block:

Prompt	Purpose
`prompts.synthesis_system`	System message for synthesis extraction
`prompts.synthesis_instructions`	Task instructions for synthesis extraction
`prompts.paper_text_description`	Description of the input paper text
`prompts.material_name_description`	Description of the target material
`prompts.synthesis_output_description`	Description of the output structure
`prompts.material_instructions`	Task instructions for material extraction
`prompts.material_input_description`	Description of material input
`prompts.material_output_description`	Description of material output

To override a prompt from the command line, wrap the whole key=value in double quotes so the shell preserves spaces. If the value itself contains a comma, add a second, inner layer of single quotes too — otherwise Hydra reads the comma as a list separator and refuses to guess which you meant:

# Focus on specific synthesis methods (no comma — plain quoting is enough)
lemat-synth extract paper.txt \
    "prompts.synthesis_instructions=Extract only sol-gel synthesis procedures. \
    Ignore characterization and testing sections."

# Customize material name handling — the value has a comma, so it needs the
# inner 'single quotes' too, or Hydra rejects it as an ambiguous list
lemat-synth extract paper.txt \
    "prompts.material_name_description='The specific compound formula to extract, \
    including all dopants and promoters.'"

Tip

Any override value with a comma needs this inner-quote treatment — "key='value, with a comma'" — not just prompt overrides. Without it, Hydra fails fast with ConfigCompositionException: Ambiguous value for argument '...'.

Managing Concurrency & Rate Limits

Two independent settings control parallel API calls:

Setting	Default	To reduce rate limits
Papers processed in parallel (batch mode)	4	`max_papers_parallel=2`
LLM calls per paper (async operations)	env-driven	`LLM_SYNTHESIS_MAX_CONCURRENT_LLM_CALLS=4` in `.env`

If you hit rate-limit errors, reduce one or both values:

# Reduce papers processed concurrently
lemat-synth batch papers/ max_papers_parallel=2

# Reduce concurrent API calls per paper
# Add to .env: LLM_SYNTHESIS_MAX_CONCURRENT_LLM_CALLS=4

Output structure

Results are written to output_dir/<paper-name>/. Each folder contains one JSON file per extracted material, plus optional performance files.

See the Output Format page for a full description of the JSON schema.

Quickstart — the shortest path to a first result
Output Format — what the result files contain
Configuration & Models — the Hydra deployment scripts, for dataset-scale and multi-LLM runs the CLI does not cover
Case Studies — domain-specific batch runs
Troubleshooting — when a run fails

CLI Reference

Quick Reference: All Arguments

Examples

Basic usage

Common customizations

Advanced: OpenRouter with multiple API keys

Advanced: Extract with performance linking (OpenRouter)

Batch Processing

Configuration Details

Model strings

API key environment variables

Allowed Environment Variables

PDF and figure processing

Domain filtering (when with_performance=true)

Batch processing options

Customizing Prompts

Managing Concurrency & Rate Limits

Output structure

Related documentation

Domain filtering (when `with_performance=true`)