Hi, I'm Vinay M Gandhi

The ingestion agent runs on APScheduler and performs change-detected upserts into MySQL, recording per-run metrics so ingestion failures surface immediately rather than silently degrading retrieval quality.

Retrieval deliberately avoids a vector store: for a corpus this structured and keyword-dense, FULLTEXT + BM25 re-ranking matched embedding-based retrieval at a fraction of the infrastructure cost — a reminder that RAG does not always require a vector database.

The full stack is FastAPI, MySQL, Jinja2, and Tailwind CSS with clean separation of concerns between ingestion, retrieval, generation, and presentation layers.

Semantic chunking uses LangChain's SemanticChunker with tuned chunk sizing and overlap management, preserving structured content — tables, images, diagrams, and architectural artifacts — as retrievable assets rather than discarding them during extraction.

Provenance tracking records source file, page number, slide reference, search score, and retrieval method for every chunk returned, so every answer can be traced back to its origin. This is what makes the system trustworthy enough for enterprise use.

Multimodal retrieval means the UI renders visual context alongside text answers, materially improving explainability for document types where a diagram carries the actual information.

Exposed via FastAPI with a Streamlit chat interface, containerised for flexible cloud or on-premises deployment.

Exceptions are retrieved from MySQL across distributed enterprise systems, normalised, and fingerprinted before any LLM call — so the expensive inference step only runs on genuinely novel failures. This is the difference between a demo and something that can run continuously against production volume.

The reporting layer generates executive-grade HTML incident reports with severity classification, confidence scores, collapsible stack traces, and syntax-highlighted code, making the output usable by management as well as engineers.