yazan-alsaleh/Multi-Agent-RAG-System

## A Multi-Agent Retrieval-Augmented Generation (RAG) System
> Upload any document and ask questions in natural language. Specialized AI agents collaborate to retrieve, reason, and return grounded answers with source citations.
## 🧩 Problem Statement
Finding specific information across large document collections is slow and error-prone. Traditional keyword search misses semantically related content, and generic chatbots hallucinate answers without grounding them in your actual documents. Existing cloud-based solutions also require sending potentially sensitive documents to third-party APIs — a dealbreaker for companies handling confidential data.
This system lets users upload any PDF or text document and ask natural language questions. A multi-agent pipeline — built with LangGraph — routes each query to specialized agents: one retrieves the most relevant chunks from a vector store, another reasons over them using a local LLM, and a critic agent verifies the answer is grounded in the source material before returning it to the user. Every answer includes the source document and the passage it came from.
The entire pipeline runs locally. No document content, query, or answer ever leaves the user's machine. The LLM (Mistral 7B via Ollama), the embedding model, and the vector store all run on local hardware — making this suitable for organizations with strict data privacy requirements such as legal, healthcare, and finance teams.
Success metric: Answers are grounded in retrieved context. Ungrounded or low-confidence answers are flagged rather than surfaced.
## Evaluation and Testing
The system was tested using machine learning / deep learning / computer vision research papers as the document corpus to evaluate document ingestion, semantic retrieval, reasoning, and answer validation.
The papers that has been used here.
The documents were ingested, chunked, embedded using Sentence Transformers, and stored in ChromaDB. Queries were then evaluated through the multi-agent pipeline consisting of retrieval, reasoning, and validation agents.
### Example Test Results
#### Attention Is All You Need Paper
Query:
> Explain transformers in LLMs and how they work?
Result:
Retrieved relevant sections from the papers.
Generated a grounded answer using Mistral 7B.
Validation Agent confirmed the answer was supported by the retrieved context.
#### YOLO Papers
Query:
> Explain how YOLO works for a beginner?
Result:
Retrieved relevant architectural information.
Reasoning Agent generated an answer grounded in the retrieved context.
Validation Agent approved the response.
#### Retrieval-Augmented Generation Paper
Query:
> What is RAG and why I need it?
Result:
Retrieved the sections explaining the motivation behind RAG.
Reasoning Agent generated an answer based only on retrieved chunks.
Validation Agent verified the answer before it was returned.
## 🏗️ Architecture
mermaid
flowchart TD
A[User] --> B[Streamlit Frontend]
B --> C[FastAPI Backend]
C --> D[LangGraph Orchestrator]
D --> E[Retriever Agent]
E --> H[(ChromaDB Vector Store)]
H --> I[Sentence Transformers Embedding Model]
E --> F[Reasoning Agent]
F --> J[Mistral 7B via Ollama]