Skip to content

RAG Configuration

Customize Retrieval-Augmented Generation (RAG) settings for improved data extraction accuracy specific to your use case.

RAG is Used by Materials Data Identifier Agent

RAG is automatically used during data extraction to retrieve relevant context from article text before querying the LLM to understand whether the article has material compositions and corresponding property values for screening. The parameters below allow you to customize RAG behavior based on your specific requirements, such as article length, domain-specific models, or extraction complexity.

Configuration Parameters

Chunking Parameters

These parameters control how articles are split into chunks for vector storage.

chunk_size (int)

Size of text chunks in characters for creating vector database embeddings.

chunk_overlap (int)

Number of overlapping characters between consecutive chunks to maintain context continuity.

Default Values

chunk_size = 1000
chunk_overlap = 25

Embedding Model Parameters

embedding_model (str)

Name of the embedding model to use for creating vector databases for RAG.

Supported Providers:

Provider Format Example
HuggingFace huggingface:model-name huggingface:thellert/physbert_cased
Sentence Transformers sentence-transformers:model-name sentence-transformers:all-mpnet-base-v2
OpenAI • openai:model-name
• model-name (default behavior)		 • openai:text-embedding-3-small
• text-embedding-3-small

Default Value

embedding_model = "huggingface:thellert/physbert_cased"

Recommended Models for Scientific Text

  • PhysBERT: huggingface:thellert/physbert_cased - A specialized text embedding model for physics, designed to improve information retrieval, citation classification, and clustering of physics literature
  • MatBERT: huggingface:pranav-s/MaterialsBERT - A fine-tuned version of PubMedBERT on a dataset of 2.4 million materials science abstracts
  • MatSciBERT: huggingface:m3rg-iitd/matscibert - A material domain language model for text mining and information extraction

Retrieval Parameters

These parameters control how relevant context is retrieved during extraction.

rag_db_path (str)

Custom path to store or load the vector databases of property-mentioned articles for RAG processing.

rag_top_k (int)

Number of most relevant text chunks to retrieve from the vector database for context.

rag_max_tokens (int)

Maximum number of tokens for RAG model responses.

Default Values

rag_db_path = "db"
rag_top_k = 3
rag_max_tokens = 512

RAG Chat Model Parameters

rag_chat_model (str)

Chat model to use for RAG-based context retrieval and synthesis.

rag_base_url (str)

Custom base URL for RAG chat model API (useful for local or custom deployments).

Default Values

rag_chat_model = "gpt-4o-mini"
rag_base_url = None

Configuration Examples

Using OpenAI

API Key Required: Set OPENAI_API_KEY in your .env file.

from comproscanner import ComProScanner

scanner = ComProScanner(output_dir="output")

# Process articles with custom chunking
scanner.process_articles(
    property_keywords={
        "exact_keywords": ["d33"],
        "substring_keywords": [" d 33 "]
    },
    rag_db_path="embeddings/piezo",
    chunk_size=800,
    chunk_overlap=50,
    embedding_model="openai:text-embedding-3-small"
)

# Extract with GPT-4o
scanner.extract_composition_property_data(
    main_extraction_keyword="d33",
    rag_db_path="embeddings/piezo",
    rag_chat_model="gpt-4o",
    rag_max_tokens=1024,
    rag_top_k=5,
)

Using Google Gemini

API Key Required: Set GEMINI_API_KEY in your .env file.

scanner.extract_composition_property_data(
    main_extraction_keyword="d33",
    rag_db_path="embeddings/piezo",
    rag_chat_model="gemini-2.0-flash",
    rag_max_tokens=1024,
    rag_top_k=4,
)

Using Anthropic Claude

API Key Required: Set ANTHROPIC_API_KEY in your .env file.

scanner.extract_composition_property_data(
    main_extraction_keyword="d33",
    rag_db_path="embeddings/piezo",
    rag_chat_model="claude-3-5-sonnet-20241022",
    rag_max_tokens=2048,
    rag_top_k=4,
)

Using Local Ollama

scanner.extract_composition_property_data(
    main_extraction_keyword="d33",
    rag_db_path="embeddings/piezo",
    rag_chat_model="ollama/llama3.1",
    rag_base_url="http://localhost:11434",
    rag_max_tokens=512,
    rag_top_k=3,
)

Using Together AI

API Key Required: Set TOGETHER_API_KEY in your .env file.

scanner.extract_composition_property_data(
    main_extraction_keyword="d33",
    rag_db_path="embeddings/piezo",
    rag_chat_model="together_ai/meta-llama/Llama-3-70b-chat-hf",
    rag_max_tokens=1024,
    rag_top_k=4,
)

Using OpenRouter

API Key Required: Set OPENROUTER_API_KEY in your .env file.

scanner.extract_composition_property_data(
    main_extraction_keyword="d33",
    rag_db_path="embeddings/piezo",
    rag_chat_model="openrouter/meta-llama/llama-3-70b-instruct",
    rag_max_tokens=1024,
    rag_top_k=4,
)

Using Cohere

API Key Required: Set COHERE_API_KEY in your .env file.

scanner.extract_composition_property_data(
    main_extraction_keyword="d33",
    rag_db_path="embeddings/piezo",
    rag_chat_model="cohere/command-r-plus",
    rag_max_tokens=1024,
    rag_top_k=4,
)

Using Fireworks AI

API Key Required: Set FIREWORKS_API_KEY in your .env file.

scanner.extract_composition_property_data(
    main_extraction_keyword="d33",
    rag_db_path="embeddings/piezo",
    rag_chat_model="fireworks_ai/accounts/fireworks/models/llama-v3-8b-instruct",
    rag_max_tokens=1024,
    rag_top_k=4,
)

Using Domain-Specific Embeddings

# PhysBERT for physics/materials science
scanner.process_articles(
    property_keywords=property_keywords,
    embedding_model="huggingface:thellert/physbert_cased",
    chunk_size=1000,
    chunk_overlap=50
)

# MatBERT for materials science
scanner.process_articles(
    property_keywords=property_keywords,
    embedding_model="huggingface:pranav-s/MaterialsBERT",
    chunk_size=1000,
    chunk_overlap=50
)

Dependencies

Install required packages based on your chosen providers:

OpenAI

pip install langchain-openai

Google Gemini

pip install langchain-google-genai

Anthropic Claude

pip install langchain-anthropic

Ollama

pip install langchain-ollama

# Install Ollama locally
# Visit: https://ollama.ai/download

Other Providers

# Together AI
pip install langchain-together

# Cohere
pip install langchain-cohere

# For HuggingFace embeddings
pip install sentence-transformers

Next Steps