Tessmora — An Omni-Modal Agentic Retrieval Platform
A self-hostable, omni-modal retrieval platform that searches documents, images, audio and video with hybrid retrieval, reranking and a bounded agent for follow-ups.
Tessmora parses documents, images, audio and video with per-modality units and dedicated vectors rather than reducing everything to plain text: a lossless agentic chunker for ordinary documents, a vision-language model plus CLIP for images, ASR plus CLAP for audio, and scene-to-shot-to-keyframe extraction for video.
Query handling combines one-pass intent detection, knowledge-base-profile routing across one, several or all knowledge bases, dense, sparse, visual, audio and video recall fused with weighted reciprocal rank fusion, and cross-encoder reranking. For complex questions, an optional bounded agent plans complementary sub-queries, retrieves them concurrently, and deduplicates evidence across rounds, using only existing read-only retrieval tools and staying within round, query and evidence-pool budgets.
Vectors are stored in Qdrant and referenced media in MinIO. It offers a full web interface, optional Feishu and Docx/Wiki import, and a built-in Codex skill and CLI, and streams answers with citations that carry source, media URL and time range.