Omni-modal retrieval, one traceable evidence chain
How can each modality keep an appropriate retrieval path and still join one inspectable answer?
Tessmora sends documents, images, audio, and video through dedicated parsing and embedding paths. One-Pass intent and knowledge-base portraits choose the search scope; dense, sparse, visual, audio, and video retrieval then feed weighted RRF, Cross-Encoder reranking, and either a direct answer or a budget-bounded read-only evidence loop.
Documents use source-faithful agentic chunks; images use VLM descriptions and CLIP; audio uses ASR and CLAP; video is indexed around semantic shots with aligned captions, speech, and key-frame visual support. The goal is to avoid collapsing every source into one text-only representation.
When a knowledge base is not pinned, topic portraits can route a query to one, several, or all relevant bases. Dense, sparse, visual, audio, and video candidates are fused by weighted RRF and narrowed by Cross-Encoder reranking.
Agent mode plans complementary subqueries and executes each round in parallel through the same read-only retriever. Round, query, and evidence budgets bound the loop; the final stream sends type-specific citations separately from answer text.
Project UI, architecture, and system visualizations used to make the work inspectable.
