How Queravel answers across documents, video, and databases
By Ekaba Bisong, SiliconBlast
Queravel answers questions across your documents, images, video, audio, databases, and cloud storage. Here is how the retrieval works.
The retrieval stack
We embed your content with E5-Mistral into 4096-dimension vectors. We store the vectors in Qdrant. We also index the text in MeiliSearch for keyword search. Two indexes, one query.
Hybrid search, then rerank
Pure vector search misses exact phrases like an error code or a part number. Pure keyword search misses meaning. We run both, merge the results, and rerank with a BGE model. You get recall and precision in one call.
Understanding video and audio
Whisper transcribes audio. Pyannote labels who spoke and when. For video, we detect scene changes, pull keyframes, and read on-screen text with OCR. A two-hour recording becomes searchable text with speakers and timestamps.
Deep research that checks itself
For hard questions, an agent plans a set of searches, runs them, verifies the sources, and writes a report. Every claim carries a numbered citation you open and read. You see where each answer came from.
What we learned
- Hybrid search beats either method alone.
- Reranking earns its cost on precision.
- Citations build trust faster than fluency.
- Async GPU workers keep long jobs from timing out.
- A polling fallback keeps the interface alive when a socket drops.
The result is one place to ask, across every source, with answers you check yourself. Bring us your data problem and we will build the same for you.