Back to Portfolio
Case Study

Support Accmcare Website

Intelligent RAG-Based Knowledge Assistant Platform with semantic search and workflow orchestration

ABOUT THIS PROJECT

Accmcare needed a full-scale RAG-powered Knowledge Assistant Platform that could deliver fast, accurate, permission-aware responses across multiple document types. The system required semantic search capabilities, vector database architecture, intelligent workflow routing, and rich chat interactions. The platform needed to handle diverse content including policies, meeting notes, Q&A, checklists, and resource documents with multi-format support (PDF, PPTX, DOCX, video transcripts).

Result:

The platform successfully delivers a fully functional, enterprise-ready Knowledge Assistant powered by state-of-the-art RAG architecture. The system provides highly accurate, context-grounded answers with seamless multi-turn interactions. The dual vector-store architecture (FAISS + OpenAI Vector Stores) enables near-instant semantic search across a large and evolving knowledge base. The intelligent chat interface with LangGraph workflow routing ensures best-fit responses every time, creating a unified interface for text, documents, and data.

OUR INVOLVEMENT

  • RAG Implementation
  • Vector DB Architecture
  • Workflow Orchestration
  • Backend Development
  • AI Integration

TECHNOLOGIES

  • OpenAI
  • LangChain
  • LangGraph
  • FAISS
  • Django
  • Python

SnakeScript's RAG-Powered Solution

We built a comprehensive RAG (Retrieval Augmented Generation) system that retrieves the most relevant information from the knowledge base before generating answers. The platform implements semantic search across diverse content types including policies, meeting notes, Q&A, checklists, and resource documents. Using LangChain-powered RAG pipeline with OpenAI embeddings, the system processes multi-format inputs (PDF, PPTX, DOCX, video transcripts) with recursive text chunking (2000 chars with 200 overlap) and metadata enrichment.

The dual vector-store architecture combines FAISS for local knowledge base indexing and OpenAI Vector Stores for assistant file attachments. The LangGraph StateGraph workflow intelligently classifies queries and routes them to appropriate handlers (RAG answers, file-based answers, or direct DB queries), ensuring conversation-aware responses that retain context across multiple messages.
The RAG system features intelligent document processing with support for multiple formats including PDF, PPTX, DOCX, video transcripts, and raw text. Each document is processed with recursive text chunking optimized at 2000 characters with 200-character overlap, ensuring comprehensive coverage while maintaining context. Metadata enrichment includes user_id, group_id, document type, and permission flags, enabling secure and permission-aware search across the knowledge base. The FAISS vector store uses OpenAI embeddings for semantic similarity search with chunk-level metadata and permission-based filtering.
Project feature

Intelligent Chat Interface & Workflow Orchestration

The platform features a smart chat system that automatically identifies query type and routes it to the appropriate handler. The LangGraph StateGraph workflow detects whether user queries relate to policies, meetings, Q&A, workflows, or generic topics, ensuring best-fit responses every time. Conversation threading with file attachments, memory retention (last 5 messages), and assistant selection creates a smooth, human-like chat experience with multi-modal support for RAG answers, file-based answers, and direct DB queries.
Multi-Modal
Response Types
Real-Time
Indexing Pipeline
Permission-Aware
Secure Search

Retrieval Augmented Generation System

Built a robust RAG system that retrieves the most relevant information from the knowledge base before generating answers.

  • 1LangChain-powered RAG pipeline with OpenAI embeddings
  • 2Recursive text chunking (2000 chars with 200 overlap)
  • 3Multi-format support: PDF, PPTX, DOCX, video transcripts, raw text
  • 4Metadata enrichment with user_id, group_id, document type, and permission flags
  • 5Conversation memory managed using ChatMessageHistory
  • 6Retrieval chain with FAISS similarity search and metadata filters

Dual Vector-Store Architecture

Implemented hybrid vector-store approach combining FAISS and OpenAI Vector Stores for optimal performance.

  • 1FAISS local file-based indexing with OpenAI embeddings (text-embedding-ada-002)
  • 2Chunk-level metadata and permission-based filtering
  • 3OpenAI Vector Stores integrated with Assistant API for file-based querying
  • 4Automatic creation of vector stores per thread/per file
  • 5Real-time indexing pipeline for new/updated documents
  • 6Interactive UI for file uploads, metadata control, and search management

Intelligent Chat & Routing System

Developed smart chat system with LangGraph workflow routing for automatic query classification and handling.

  • 1LangGraph StateGraph workflow for query type detection
  • 2Automatic routing to RAG, file-based, or direct DB query handlers
  • 3Conversation threading with file attachments and memory retention
  • 4Thread-level persistence of last 5 messages
  • 5Auto-selection of the right model/assistant based on context
  • 6Multi-modal support: RAG answers, file-based answers, and direct DB queries

Key Results

1

Highly Accurate, Context-Grounded Answers

RAG system delivers precise responses by retrieving relevant information before generation

2

Near-Instant Semantic Search

Dual vector-store architecture enables fast search across large and evolving knowledge base

3

Intelligent Workflow Routing

LangGraph automatically classifies queries and routes to best-fit handler for optimal responses

4

Enterprise-Ready Platform

Fully functional, scalable solution with permission-aware search and multi-modal support

Success Story

This project delivered a fully functional, enterprise-ready Knowledge Assistant powered by state-of-the-art RAG architecture, intelligent routing, and multi-modal chat. The platform ensures accuracy, security, and scalability with seamless multi-turn interactions, permission-aware search, and consolidated access across policies, meetings, and resource documents. The hybrid vector-store approach and optimized chunking strategy make it suitable for growing enterprises.

Read the full success story: https://docs.google.com/document/d/1l0t0MVAYZWij_vjqjS5UieVG3bC-ba30NE9WXEApIUs/edit?usp=sharing