or-lib 1.3.0

OptimisedRAG: Simple and Fast Retrieval-Augmented Generation a modified version of LightRAG

or-lib

or-lib is a modular extension of LightRAG that enriches the Retrieval-Augmented Generation (RAG) pipeline by integrating graph-based algorithms for efficient and quality-enhanced retrieval and image-based query support for multimodal reasoning. It builds upon LightRAG’s hybrid architecture to improve both retrieval accuracy and user interactivity. Enhanced by Md Nazish Arman

---

🔍 Key Enhancements Over LightRAG

1. Graph-Based Retrieval Optimization

Introduces several graph algorithms to rank and filter knowledge graph nodes and relationships for more relevant information retrieval:

• Degree Centrality
• PageRank
• Article Rank (personalized PageRank)
• Betweenness Centrality
• CELF-Based Influence Maximization

These algorithms help dynamically identify high-impact entities and relations in the knowledge graph, improving time by 50% and retrieval quality by 30%.

2. Image Query Support

Enhances RAG to handle image-based prompts via pre-indexed image metadata and summaries:

• Extracts and processes image chunks in _build_query_context()
• Associates each image with a unique S3 image_id
• Returns presigned image URLs for downstream consumption
• Adds visual context understanding into RAG flows

---

📦 Features

Feature	Description
GraphAlgorithms	Modular class with pluggable centrality/influence metrics. Used at query time to score entities.
QueryParam.graph_algorithm	Users can dynamically select which graph algorithm to apply per query (pagerank, degree_centrality, etc).
Image Chunk Processing	Enhances query results with structured image summaries and image IDs that can be mapped to S3-hosted images.
Presigned URL Integration	Supports image result delivery through S3-backed URL mapping.

---

🧠 Graph Algorithm

The GraphAlgorithms class (in algorithms.py) provides the following methods:

compute_degree_centrality(node_datas, edge_datas, k, weighted=False)
compute_pagerank(node_datas, edge_datas, k)
compute_article_rank(node_datas, edge_datas, k)
compute_betweenness_centrality(node_datas, edge_datas, k)
celf_influence_maximization(node_datas, edge_datas, k)

Each method updates node_datas with rank scores and returns the top k nodes.

---

🧾 Usage Example

from orlib.algorithms import GraphAlgorithms
graph_algo = GraphAlgorithms()

top_nodes = graph_algo.compute_pagerank(node_datas, edge_datas, k=10)

To use it in a query:

query_param.graph_algorithm = "pagerank"
response, image_ids = await kg_query(
    query,
    knowledge_graph_inst,
    entities_vdb,
    relationships_vdb,
    text_chunks_db,
    query_param,
    global_config
)

---

🖼️ Image Support

✨ What It Does:

• Parses uploaded documents and stores image summaries as chunks, along with relevant metadata such as image_id (used for S3 mapping).
• During image-related queries, retrieves the relevant image chunks.
• Extracts summaries and metadata for matching image chunks.
• Sends the image summaries in CSV format to the LLM.
• Filters out image IDs whose summaries are most relevant to the query.
• Implements caching of image IDs related to previous queries to avoid redundant processing.
• Returns the relevant image IDs.
• Generates presigned URLs for image access.

🔁 Flow:

1. Image metadata is indexed with image_id and content.

2. During a query:

   • Keywords are matched against stored image chunks.
   • Relevant results are structured into an image_chunk.

3. An img_prompt is generated using:

   image_csv_data = "serial number,image id,image summary\n1,img001,..."
   

4. The LLM receives both textual and visual context for improved relevance.

🗂️ Image Storage

Images are expected to be pre-processed and stored in S3. The corresponding image_id is then used to generate presigned URLs for secure frontend rendering.

---

🧪 QueryParam Extensions

QueryParam(
    graph_algorithm="pagerank",
    only_need_prompt=False,
    top_k=60,
    response_type="Bullet Points",
    ...
)

• graph_algorithm: Selects the algorithm to guide ranking logic in the retrieval phase.
• top_k: Defines how many top nodes or relationships to consider.
• only_need_context / only_need_prompt: Controls which intermediate step to return (useful for debugging or chaining outputs).

---

✅ Supported Graph Algorithms

Algorithm	Purpose
pagerank	Scores nodes based on importance across the graph
degree_centrality	Scores nodes by connection count
article_rank	Personalized PageRank for localized influence
betweenness_centrality	Captures bridge nodes that connect clusters
celf_influence	Approximates influence spread using CELF optimization

---

📌 Requirements

• Python 3.10+
• networkx
• LightRAG dependencies (faiss, transformers, langchain, etc.)
• boto3 or any S3-compatible client for presigned URLs

---

Author

Md Nazish Arman

• 🌐 GitHub
• 💼 LinkedIn