Request for Feedback on LazyGraphRAG Implementation

[sagarpurohit88]

Hello Everyone,

I am currently working on building a LazyGraphRAG system for internal documentation.

Problem Statement:
Although the LazyGraphRAG design is intended to be dynamic and efficient, the initial query retrieval is taking approximately 22 minutes. This time includes generating embeddings, constructing communities, and extracting relationships. For subsequent (follow-up) queries, the retrieval time is reduced to around 1 minute, since only embeddings need to be built.

I would like to better understand:

• How we can reduce the overall retrieval time, especially for the initial query.
• Whether this is considered the correct or optimal approach for implementing LazyGraphRAG.
• If there are any reference implementations, best practices, or related resources that I could explore.

For context, I have included below a detailed breakdown of the complete LazyGraphRAG flow, outlining each component and how they interact within the system.

Stage 1

• Sets up global configuration variables
• Initializes Neo4j driver connection
• Configures limits for different retrieval strategies
• No graph building happens here - just setup

Stage 2

• Connects to Confluence API
• Fetches 9 specific pages about specific platform topic
• Cleans HTML content using removing tag.
• Splits documents into 512-token chunks with 20-token overlap
• Creates LangChain Document objects with metadata

Stage 3

• Clears existing Neo4j data
• For each document chunk:
• Extracts named entities using spaCy NLP
• Creates Chunk node in Neo4j with text + metadata
• Creates Entity nodes for extracted entities
• Creates MENTIONS relationships: (Chunk)-[:MENTIONS]->(Entity)

Stage 4

Query Processing Flow

The system runs three parallel retrieval strategies:

A. Semantic Retrieval

• Lazy loads embeddings if not present
• Embeds the query
• Calculates cosine similarity with chunk embeddings
• Returns most similar chunks

B. Graph Traversal

• Extracts entities from query using spaCy
• Finds chunks that mention these entities directly
• Traverses 1-2 hops via RELATED relationships
• Returns related chunks

C. Community Analysis

• Lazy builds communities if not present
• Embeds query and compares with community summaries
• Returns relevant community summaries for global context

Result Fusion

• Combines all three retrieval results
• Deduplicates based on content
• Ranks by similarity scores
• Limits to configured maximum results

Answer Generation

1. Formats context from all sources with source labels
2. Uses enhanced prompt that instructs GPT-4:

• Synthesize detailed chunks + high-level summaries
• Identify entity relationships
• Connect concepts across the knowledge graph

3. Generates comprehensive answer

Complete Data Flow Example