Brain Search
Find any file by describing it. Without sending a single byte to the cloud.
Brain Search indexes your PDFs, Word docs, spreadsheets and notes on your own machine, then uses on-device AI to understand what you mean, not just what you typed.
C# and .NET 10, with an Avalonia desktop UI and an embedded ASP.NET Core engine. Meaning comes from on-device embeddings (ONNX Runtime, bge-small-en). Keyword ranking is Lucene.NET (BM25). Dense vectors live in SQLite via sqlite-vec, and the two result sets are merged with reciprocal rank fusion (RRF).
Why teams choose it
01 · PRIVACY
Your files never leave your machine. Not as cache, not as embeddings, not as metadata. Brain Search is fully on-device, which is what makes it safe to deploy inside a company.
02 · INTENT
Search by what you mean, not what you remember. Ask for the warehouse-lease contract and find it, even when the file is named IMG_2841.pdf. Hybrid keyword plus semantic ranking does the work.
03 · LOCAL
Built on Lucene.NET for keyword search and on-device ONNX embeddings for meaning, fused with RRF. No API keys, no cloud roundtrips, no vendor lock-in. It runs offline.
brain_search · on-device
LUCENE.NET · BM25
- lease-terms-2024.pdf11.4
- IMG_2841.pdf9.7
- rent-schedule.xlsx6.2
ONNX · bge-small-en
- IMG_2841.pdf0.91
- warehouse-floorplan.pptx0.78
- lease-terms-2024.pdf0.74
↑ 8,432 files indexed locally · 12.3 GB · 1.2s
keyword + semantic, fused with RRF, all on-device
How it works
Point Brain Search at a folder. It parses every file (PDF, DOCX, PPTX, XLSX, HTML, Markdown, plain text and email), splits the text into chunks, and embeds each chunk with a local ONNX model. When you search, Lucene.NET handles exact keyword matching and sqlite-vec handles meaning; reciprocal rank fusion blends both into one ranked list, grouped by file. The model and the index both live on your machine. There is no cloud component.
Supported formats
- DOCX
- PPTX
- XLSX
- HTML
- Markdown
- TXT
Bring it to your team.
Brain Search is built for teams that need internal document search without sending anything to a third party. Three deployment tiers (Solo, Team, Enterprise) cover individuals through on-prem, sovereign installs. Conservative modelling puts the saving at two to three hours per employee per day versus filename-and-folder hunting.