graphrag prompt-tune

Generate custom prompts tuned to your domain and data

The prompt-tune command automatically generates domain-specific prompts by analyzing your input documents. This creates prompts that are better suited to your specific use case than the default generic prompts.

Usage

graphrag prompt-tune [OPTIONS]

Options

--root · string · default: "current directory"

The project root directory.

Aliases: -r

--domain · string

The domain your input data is related to. For example: “space science”, “microbiology”, “environmental news”, “legal documents”.

If not specified, the domain will be automatically inferred from the input data.

--output · string · default: "prompts"

The directory to save generated prompts to, relative to the project root directory.

Aliases: -o

--selection-method · string · default: "random"

The text chunk selection method for prompt generation.

Available methods:

• random - Randomly select chunks (fastest)
• top - Select top chunks by order
• auto - Automatically select diverse chunks using embeddings (best quality)

--limit · integer · default: "15"

The number of documents to load when --selection-method is random or top.

--n-subset-max · integer · default: "300"

The number of text chunks to embed when --selection-method=auto. Higher values provide more diversity but take longer.

--k · integer · default: "15"

The maximum number of documents to select from each centroid when --selection-method=auto.

--max-tokens · integer · default: "2000"

The maximum token count for prompt generation examples.

--min-examples-required · integer · default: "2"

The minimum number of examples to generate/include in the entity extraction prompt.

--chunk-size · integer · default: "1200"

The size of each example text chunk in tokens. Overrides chunking.size in the configuration file.

--overlap · integer · default: "100"

The overlap size for chunking documents in tokens. Overrides chunking.overlap in the configuration file.

--language · string

The primary language used for inputs and outputs in GraphRAG prompts.

Examples: “English”, “Spanish”, “French”, “Japanese”

If not specified, defaults to English.

--discover-entity-types · boolean · default: "true"

Discover and extract unspecified entity types from the data.

Use --no-discover-entity-types to disable and only use predefined entity types.

--verbose · boolean · default: "false"

Run the prompt tuning pipeline with verbose logging.

Aliases: -v

Examples

Basic prompt tuning

Generate prompts with automatic domain inference:

graphrag prompt-tune

Specify domain

graphrag prompt-tune --domain "medical research"

Use auto selection for better quality

graphrag prompt-tune --selection-method auto

The auto selection method uses embeddings to select diverse, representative samples from your documents.

Customize output directory

graphrag prompt-tune --output custom_prompts

Multilingual prompt generation

graphrag prompt-tune --language "Spanish" --domain "noticias políticas"

Increase sample diversity

graphrag prompt-tune \
  --selection-method auto \
  --n-subset-max 500 \
  --k 20

Use predefined entity types only

graphrag prompt-tune --no-discover-entity-types

Custom chunking parameters

graphrag prompt-tune \
  --chunk-size 800 \
  --overlap 200

Verbose output

graphrag prompt-tune --verbose

Generated prompts

The prompt-tune command generates three customized prompt files:

extract_graph.txt

Prompt for extracting entities and relationships from text. This is customized with:

• Domain-specific entity types discovered from your data
• Example extractions from your actual documents
• Domain context and terminology

summarize_descriptions.txt

Prompt for summarizing entity descriptions. Customized with:

• Domain-specific summarization guidelines
• Examples from your data

community_report.txt

Prompt for generating community reports. Customized with:

• Domain-appropriate report structure
• Relevant analysis dimensions for your domain

How it works

1. Sample selection: Selects representative text chunks from your documents using the specified selection method
2. Domain analysis: Analyzes the samples to understand domain characteristics and terminology
3. Entity discovery: Identifies domain-specific entity types present in your data
4. Example generation: Creates few-shot examples by extracting entities from sample chunks
5. Prompt creation: Generates prompts incorporating discovered entity types and examples

Selection methods

Random selection

• Fastest method
• Randomly samples chunks from documents
• Good for homogeneous datasets
• Use --limit to control sample size

Top selection

• Selects first N chunks in order
• Fast and deterministic
• Good when early content is representative

Auto selection (recommended)

• Uses embeddings to find diverse, representative samples
• Creates embeddings for up to --n-subset-max chunks
• Clusters chunks and selects up to --k from each cluster
• Best quality but slower due to embedding generation
• Recommended for heterogeneous datasets

Best practices

1. Use auto selection: For best results, use --selection-method auto with adequate sample size
2. Specify domain: Providing a domain helps generate more focused prompts
3. Diverse samples: Ensure your input documents are representative of your full dataset
4. Review outputs: Always review generated prompts before using them for indexing
5. Iterate: Try different selection methods and parameters to find what works best

Customizing generated prompts

After generation, you can manually edit the prompts in the output directory:

1. Review the generated entity types and examples
2. Add or remove entity types as needed
3. Adjust the prompt instructions for your specific needs
4. Test with a small index before running on full dataset

Using tuned prompts

After generating prompts:

1. Review the prompts in the prompts/ directory (or your custom output directory)
2. Copy them to your project’s prompts/ directory if using a custom output location
3. Run indexing - it will automatically use the new prompts:
   graphrag index
   

Performance considerations

• Auto selection: Slower due to embedding generation but produces better quality
• Sample size: Larger --n-subset-max values take longer but capture more diversity
• Token limits: Higher --max-tokens allows for more examples but increases API costs

Error handling

Common issues:

• No input documents: Ensure documents exist in your input/ directory
• API errors: Check that your API key is configured in .env
• Insufficient samples: Increase --limit or --n-subset-max if you get too few examples
• Language mismatch: Ensure --language matches your document language

Next steps

• Review and customize generated prompts
• Run indexing with tuned prompts
• Understand entity types
• Configure prompt settings