Production RAG assistant that ingests company documents and answers questions with source citations. Custom retrieval pipeline with vector search, chunking strategies, and streaming LLM responses.
10K+ documents indexed
<800ms time-to-first-token
95% citation accuracy
Companies struggle to make internal knowledge accessible. Employees waste hours searching through scattered docs, wikis, and Slack threads to find answers.
Document ingestion pipeline with PDF/DOCX/TXT parsing
Custom chunking strategy with overlap for context preservation
OpenAI embeddings → Pinecone vector store for semantic retrieval
Hybrid search combining keyword matching + vector similarity
Streaming LLM responses with source citation linking
Next.js frontend with real-time chat interface
Optimized chunking strategy to preserve context across document sections - tested 5 approaches before settling on recursive splitting with 200-token overlap
Built citation system that maps LLM output spans back to source documents with paragraph-level precision
Achieved sub-second time-to-first-token through streaming and parallel retrieval