RAG Comparative POC for Knowledge Base

@divyam12345

Dear, I have extensive experience designing and implementing LLM-powered RAG systems, including vector-based retrieval, knowledge graph pipelines (Neo4j/Graph RAG), and hierarchical document retrieval frameworks. I’ve previously built evaluation-driven GenAI systems where fair benchmarking across multiple retrieval strategies was critical, so I fully understand that this POC is not just about implementation—but about controlled comparison, reproducibility, and insight generation. I will also design a robust evidence normalization layer and evaluation harness to compare accuracy, grounding, citation quality, and latency in a structured way. My proposed stack includes Python + LlamaIndex/LangChain (custom where needed), FAISS or Pinecone for vector RAG, Neo4j for Graph RAG, and structured document parsing (PyMuPDF + heading detection) for Tree RAG, with OpenAI/Azure OpenAI for generation. I will implement a clean modular pipeline covering ingestion, indexing, routing, retrieval, normalization, and evaluation—along with logging and reproducibility controls. The final deliverables will include a working codebase, comparison dashboard/report, and a clear recommendation backed by empirical results, aligned with your success criteria.

@niveditaj1

Hello, I will design a modular evaluation framework to ingest your source documents and build the three distinct indexing structures. I will implement a standard vector database for the normal RAG, a graph-based retrieval system for entity relationships, and a hierarchical tree structure for section-based data. I will then develop an automated testing script to run the same set of questions against all three engines, capturing performance metrics and accuracy. Finally, I will provide a detailed comparison report to help your orchestrator dynamically select the best retrieval strategy based on the query type. This approach ensures a fair comparison and a clear path toward building an intelligent routing layer for your agent. 1) What is the approximate size and format of your current knowledge base documents? 2) Do you have a preferred LLM provider or a specific cloud environment for the POC? 3) Are there specific evaluation metrics, such as latency or precision, you want me to prioritize? Thanks, Nivedita

@indreshpnk

Hello, I understand you need to build a POC comparing Normal RAG, Graph RAG, and Tree RAG on the same enterprise knowledge base. The goal is to deliver a reliable, evaluation-driven solution that fairly benchmarks all approaches and provides clear recommendations. Here’s what I can provide: End-to-end pipeline to ingest documents and build vector, graph, and hierarchical indexes with proper metadata Orchestrator with question classification and dynamic routing for runtime and comparison modes Evaluation harness to run all approaches on the same question set and generate a detailed comparison report I bring over 4+ years of experience in Python, LLMs, LangChain, and vector/graph databases, with a strong focus on scalable and evaluation-driven AI systems. I’ve worked on RAG pipelines, document intelligence, and citation-grounded QA systems. Just to clarify a few things: Will you provide the initial knowledge base and labeled question set for evaluation? Do you have any preference for vector DB or graph DB like Neo4j? Please come to the chat box to discuss more about your project. Best regards Indresh Kushwaha

@letswork92

I understand you're aiming to evaluate the effectiveness of different retrieval approaches for your enterprise knowledge base. This is a complex but exciting challenge that requires a structured approach to ensure fair comparisons between Normal RAG, Graph RAG, and Tree RAG. With over 12 years of experience in AI/LLM systems, I have developed similar proofs of concept using technologies like Python, Neo4j for graph databases, and various vector databases for embedding and indexing. For implementation, I would create distinct ingestion pipelines for each retrieval method while ensuring all outputs are normalized against a common evidence schema. To ensure a fair comparison, I'll maintain consistent evaluation metrics across all approaches and leverage my expertise in prompt design to generate reliable answers from the same foundation model. My estimated timeline is about 8 weeks, with an approximate budget of $25k. Key risks include data quality issues and potential discrepancies in retrieval efficiency based on document structure. What specific types of questions do you envision using for the evaluation phase?

@krishnap467

• Languages: streamlit • Tools: • Data Tool: Python, C++, JavaScript, SQL, C, Oops, DSA, Html, css, Numpy, padas , Matplotlib, seaborn, vscode, Jupyter,colab, PostgreSQL, MySQL Excel, Power BI • Data Analysis: Data Cleaning, Data Visualization • Statistics: Mean, Median, Correlation, Hypothesis Basics

@cyberfabrica

Hello there, I’ve carefully read your project details and this is exactly the kind of structured RAG evaluation system I specialize in. Approach & Stack: Python + LlamaIndex/LangChain, FAISS/Weaviate (Vector RAG), Neo4j (Graph RAG), and a custom hierarchical parser for Tree RAG. LLM via OpenAI/Azure. Modular pipeline to ensure identical ingestion, prompting, and evaluation across all approaches. Implementation: Build 3 parallel indexes from same documents (vector, graph with entity-linking, tree hierarchy with structure audit) Orchestrator for routing + separate comparison mode Common evidence schema for fair grounding Fixed prompt + model settings across all runs Evaluation harness with ~40 tagged questions + scoring (accuracy, faithfulness, latency, etc.) Fair Comparison: Strict normalization, identical prompts, same LLM, same question set, and logged metrics for side-by-side analysis. Timeline: 2–4 weeks Budget: 100000 INR (POC scope) Deliverables: full codebase, ingestion pipelines, APIs, evaluation dashboard/report, and final recommendation with clear trade-offs. Why choose me: deep experience in RAG systems, evaluation frameworks, and scalable AI architecture with strong focus on correctness over hype. Regards, Raakesh CyberFecbrica Software Solutions

@dhirajg32

I have an working experience in this domain for more than 6 years. I have a very good knowledge in this area.

@ghkjois

Hi, this is a well-defined RAG comparison POC and I’ve already worked on similar systems involving vector RAG, graph-based retrieval, and evaluation frameworks; I can build all three pipelines (Normal, Graph, Tree RAG) on the same dataset with a unified evidence schema, fair comparison harness, and a routing layer for strategy selection, along with detailed logging and a final evaluation report; I propose using Python with Chroma, numpy, pandas, pyMuPDF, tqdm, sentence-transformers, tabulate, and OpenAI/Anthropic LLMs; estimated timeline is 7–10 days with a budget of ₹10,000, assuming one primary KB and minimal UI, and I will ensure the comparison is data-driven and properly documented. Ping for output snapshots.

@kosei4

Dear Client, Thank you for posting this project. I would be truly grateful for the opportunity to work on this RAG comparison POC. Your specification is one of the most thorough briefs I have seen, and I would be honored to bring it to life. I build AI-powered automation systems in production daily. I run a 14-step automated workflow on Cloudflare Workers integrating OpenAI/Claude APIs, Slack API, and Gmail API with full logging. My stack: Python 3.11 + LangChain + LlamaIndex. Normal RAG with ChromaDB/FAISS. Graph RAG with Neo4j. Tree RAG with custom hierarchical parser. GPT-4o as foundation model (same config across all 3). Automated evaluation harness with comparison matrix. My approach: document structure audit, common evidence schema, same FM/prompt/temperature, full logging, 30-50 tagged questions compared side-by-side. POC correctness matters more than polish. Clean documented code with setup instructions and data-backed recommendation report. Message-only communication. No calls needed. I would deeply appreciate the chance to prove my capabilities. Respectfully, Kosei Izumida