SQL processing over images, text, and audio (itrummer.github.io)

ThalamusDB is a DBMS that lets you run SQL-style queries over multimodal data by embedding LLM-powered semantic operators directly into SQL. Example: SELECT on HolidayPictures with NLFILTER(H.pic, 'this is a picture of the beach') plus NLJOIN(H.pic, P.pic, 'the same person appears in both pictures') finds beach photos that show Alice or Bob. The system accepts standard SQL types and treats file paths stored in text columns as first-class objects: it auto-detects PNG/JPG/JPEG images and WAV/MP3 audio, picks suitable LLMs to analyze them, and exposes NLFILTER/NLJOIN-style predicates that operate on text, images, and audio. What sets ThalamusDB apart is its emphasis on approximate query processing to control cost and latency. Users can set per-query cost bounds and error constraints; the engine processes data subsets, returns progressive partial results, and computes deterministic bounds on aggregates and answers while minimizing LLM calls. That design makes semantic, multimodal querying practical at scale for dataset curation, retrieval, analytics, and human-in-the-loop workflows. Compared to other systems (e.g., LOTUS, Palimpsest, FlockMTL, CAESURA), ThalamusDB uniquely combines audio support and a semantic-SQL interface with explicit cost/error controls for bounding expensive LLM-driven operations.