Bypassing the Multimodal Tax: Framework-Free Hybrid RAG, Raw SQL RRF, and Live UI Telemetry

A framework-free approach to production RAG: parsing documents locally to avoid the "multimodal tax" of vision-API costs, running hybrid dense + sparse retrieval directly in PostgreSQL with Reciprocal Rank Fusion (no vector database, no LangChain), and wiring up full observability with Langfuse. Covers the schema, the raw SQL, agent mode vs. direct RAG, upload and query-time guardrails, and the economics of architecture-first optimization over prompt engineering. Fully reproducible via GitHub Codespaces.
Originally presented at AI Engineer World's Fair 2026, Online Track.

Transcript