RAG Agent

Expert agent for retrieving and synthesizing information from scientific literature using Retrieval-Augmented Generation.

Table of Contents

• Overview
• Capabilities
• How It Works
• Example Prompts
• Configuration
• Troubleshooting
• See Also

Overview

The RAG Agent specializes in searching and synthesizing information from a knowledge base of scientific papers on HEM research. It uses vector similarity search to find relevant documents and generates comprehensive answers with proper citations.

Expertise Areas

• Scientific literature retrieval
• HEM and AEM research papers
• Cation design literature
• Alkaline stability studies
• Membrane property research
• Citation and source management

Backend

The RAG Agent uses Qdrant vector database for document storage and retrieval, with embeddings generated by a configured embedding model.

Capabilities

Capability	Description
Document Retrieval	Find relevant papers based on semantic similarity
Context Synthesis	Combine information from multiple sources
Citation Management	Track and cite sources properly
Knowledge Gap Identification	Identify areas lacking research
Research Direction Suggestions	Propose relevant research directions

How It Works

Architecture

graph TD
    A[User Query] --> B[RAG Agent]
    B --> C[HEMRetriever]
    C --> D[Embedding Model]
    D --> E[Query Vector]
    E --> F[Qdrant Search]
    F --> G[Retrieved Documents]
    G --> H[Context Building]
    H --> I[LLM Synthesis]
    I --> J[Cited Response]

Retrieval Process

1. Query Processing: User query is converted to an embedding vector
2. Similarity Search: Qdrant finds the k most similar document chunks
3. Context Building: Retrieved documents are formatted with metadata
4. Synthesis: LLM generates a response using the retrieved context
5. Citation: Sources are cited using [Source X] Author et al. (Year) format

HEMRetriever

The HEMRetriever class handles document retrieval:

class HEMRetriever:
    def retrieve_with_context(self, query: str, k: int = 5) -> dict:
        """
        Retrieve relevant documents with context.
        
        Returns:
            {
                'context': str,      # Formatted context for LLM
                'sources': list,     # Source metadata
                'num_sources': int   # Number of sources found
            }
        """

Response Format

The RAG Agent generates responses with:

• Synthesized information from multiple sources
• Proper citations: [Source X] Author et al. (Year)
• Source count summary at the end

Example Prompts

Literature-Driven Design

Retrieve recent literature on cation designs for hydroxide exchange membranes 
with high alkaline stability. Summarize typical structural motifs and then 
propose 5 new candidate cations that follow those design principles.

Mechanism and Design Loop

Explain the main degradation mechanisms for quaternary ammonium cations in 
AEMs under alkaline conditions. Then propose new cation designs that mitigate 
these mechanisms, and evaluate them using your available tools.

Research Survey

What does the literature say about the relationship between cation structure 
and ionic conductivity in anion exchange membranes? Cite the relevant studies.

Gap Analysis

Based on the available literature, what are the main knowledge gaps in 
understanding alkaline stability of imidazolium-based cations? 
Suggest research directions to address these gaps.

Comparative Analysis

Compare the reported alkaline stability of piperidinium vs tetraalkylammonium 
cations based on the scientific literature. Which structural features 
contribute to better stability?

Configuration

Environment Variables

Variable	Purpose	Default
QDRANT_URL	Qdrant server URL	http://localhost:6333
QDRANT_API_KEY	Qdrant API key (if required)	None
EMBEDDING_MODEL	Embedding model name	Configured in settings

Qdrant Setup

The RAG system requires a running Qdrant instance with indexed documents:

# Start Qdrant (Docker)
docker run -p 6333:6333 qdrant/qdrant

# Or use the embedded Qdrant storage
# Located at: OHMind_agent/qdrant_storage/

Document Ingestion

Scientific papers are ingested using the ingestion script:

python OHMind_agent/scripts/ingest_papers.py

Documents are stored in:

OHMind_agent/data/scientific_papers/

Embedding Configuration

The embedding model is configured in OHMind_agent/rag/embeddings.py:

# Typical configuration
embedding_model = "text-embedding-ada-002"  # OpenAI
# or
embedding_model = "BAAI/bge-large-en"  # Local model

Troubleshooting

Common Issues

“Embeddings not configured”

Cause: The embedding service is not available.

Solutions:

1. Check that the embedding API endpoint is accessible
2. Verify API credentials are set correctly
3. Ensure the vector database is initialized

No Results Found

Cause: Query doesn’t match indexed documents.

Solutions:

1. Try rephrasing the query
2. Use more specific HEM-related terminology
3. Check that documents have been ingested

Connection Errors

Cause: Qdrant server not running.

Solutions:

1. Start the Qdrant server
2. Check QDRANT_URL environment variable
3. Verify network connectivity

Error Messages

The RAG Agent provides helpful error messages:

# Example error handling
if "Embeddings not configured" in error_msg:
    user_msg = (
        "I couldn't search the literature database because the embedding "
        "service is not available. This could be due to:\n"
        "- The embedding API endpoint is not accessible\n"
        "- Missing or invalid API credentials\n"
        "- The vector database is not initialized\n\n"
        "Please check your configuration and try again."
    )

Data Sources

Indexed Content

The RAG system indexes scientific papers related to:

• Hydroxide exchange membranes (HEM)
• Anion exchange membranes (AEM)
• Cation design and synthesis
• Alkaline stability studies
• Ionic conductivity research
• Polymer membrane properties

Document Format

Documents are chunked and indexed with metadata:

{
    "content": "Document text chunk...",
    "metadata": {
        "title": "Paper Title",
        "authors": "Author et al.",
        "year": 2023,
        "doi": "10.1234/example",
        "source": "filename.pdf"
    }
}

See Also

• Agent Reference - Overview of all agents
• Web Search Agent - For real-time web information
• Literature Search Tutorial - Step-by-step guide
• Configuration Reference - Qdrant setup

---

Last updated: 2025-12-22 | OHMind v1.0.0