Upload PDFs, Word documents, presentations, images, audio and video, then ask questions in plain language. ilm answers with citations that open the exact page, slide, image or second of video the claim came from. It runs fully locally.

Every source becomes a segment (a PDF paragraph, a rendered page, a video scene or a transcript window) with a locator, so retrieval and citation work the same way across all modalities. Three retrievers run in parallel: BM25 for exact identifiers, BGE-M3 dense embeddings for paraphrase, and SigLIP 2 cross-modal search for things nobody described in writing. Their results are fused by reciprocal rank and reranked.

Multi-part questions are decomposed by a planner, each part is retrieved separately and the evidence is merged, which fixed cases where a single embedding landed between two topics and retrieved neither. Without an LLM it answers extractively and cannot hallucinate; with any OpenAI-compatible endpoint (vLLM) it writes grounded, cited answers. FastAPI backend with a Next.js frontend.