RAG Expert AI ML

Project Overview We are building a domain-specific RAG (Retrieval-Augmented Generation) chatbot designed to answer complex, high-accuracy queries related to GST rates and legal compliance. Our knowledge base is derived from official CBIC PDFs containing narrative legal text and highly structured rate tables. The primary challenge is maintaining "ground truth" across frequently updated legal documents and ensuring perfect data extraction. Key Responsibilities Advanced Table Extraction: Develop robust, automated pipelines to extract and normalize complex rate tables from multi-page PDFs, ensuring structural integrity where headers and layouts vary. Retrieval Optimization: Refine hybrid search strategies (dense/sparse) to bridge the gap between user terminology ("chappals") and legal taxonomy ("footwear of sale value..."). Version Control for LLMs: Architect a solution for handling "updated info" (amendments, corrigenda, and superseding notifications) to ensure the model always pulls the current version of the truth, rather than superseded data. Deterministic Retrieval: Implement precision-first retrieval mechanisms to support exact-match requirements (e.g., specific HSN codes or tax rates), reducing reliance on approximate vector similarity where absolute accuracy is mandatory. Technical Requirements Deep Experience with RAG: Proven track record in building RAG systems for domain-specific, high-accuracy environments. PDF Parsing Proficiency: Strong experience with tools like Docling, Unstructured, or custom parsing logic to handle complex, multi-year PDF layouts. Semantic Search & Reranking: Expertise in hybrid search (BGE-M3, Qdrant) and implementing reranking pipelines to improve retrieval precision. Stack: Backend: FastAPI. Search/Database: Qdrant (Hybrid search), PostgreSQL. Models: Experience working with local SLMs (e.g., Qwen 3.5) for both reasoning and ingestion-time enrichment. Data Engineering: Understanding of document lifecycle management (linking amendments to original sources).