AI Search

Fluent Search includes a local AI-powered semantic search engine. When enabled, search results are ranked not just by keyword matching, but by meaning — so a search for "picture editor" can find "Photoshop," and "messaging app" can find "Discord."

All AI processing runs locally on your device. No data is sent to external servers.

Setup

Download the AI model

  1. Go to Settings → AI Settings.
  2. Click Manage AI components → Download model.
  3. Choose a model:
Model Size Languages
English Only ~32 MB English
Multi Language ~117 MB 100+ languages (Chinese, Japanese, Arabic, Hindi, European languages, and more)

The download includes both the language model and the ONNX runtime. Files are stored in %AppData%\Blast\AI\Embeddings\.

After downloading, the model loads automatically. You can delete it later from the same button to free disk space.

How it works

Passive mode (default)

When AI search is enabled but the AI search tag is not active, Fluent Search operates in cached-only mode:

  • Results that have been previously embedded (cached) get AI-boosted ranking
  • No new AI inference runs during search — this keeps search instant
  • The cache is pre-warmed with search tag names on startup

Active mode (AI tag)

When you activate the AI search tag (type AI → Tab), Fluent Search switches to full AI mode:

  • Embeddings are generated on-the-fly for all results
  • Search delay is slightly increased to allow inference to complete
  • Results are ranked by semantic similarity to your query

Scoring

AI similarity scoring uses cosine similarity between your query and each result's text:

  • AI scores are always below perfect text-match scores — exact keyword matches still rank highest
  • A minimum similarity threshold (default 0.6) filters out weak matches
  • This means AI enhances results without overriding strong text matches

Summarize

When the AI model is loaded, a Summarize operation (✨) appears on file and browser tab results.

How it works:

  1. Select a file or browser tab result
  2. Click the Summarize operation
  3. Fluent Search extracts the text content:
    • Files: Reads text from documents (supports Office, PDF, plain text via Windows IFilter)
    • Web pages: Downloads and extracts the article text
  4. The AI selects the 6 most representative sentences using Maximal Marginal Relevance
  5. A summary panel appears with bullet-pointed key sentences

Summarization is extractive — it selects existing sentences rather than generating new text, so summaries are always faithful to the source.

GPU acceleration

By default, AI inference runs on your CPU. You can enable GPU acceleration in Settings:

Setting Description
Use GPU Uses DirectML (works on any DirectX 12 GPU — NVIDIA, AMD, Intel)

GPU performance varies by hardware. On some systems it's faster; on others, CPU may be better. If GPU inference fails, Fluent Search automatically falls back to CPU.

The AI model runs in a separate process to isolate it from the main application — GPU driver issues won't crash Fluent Search.

Settings

Setting Default Description
Disable all AI features Off Completely turns off AI and hides the AI search tag
Enable AI search On Use AI for semantic result ranking
Use GPU Off Run inference on GPU instead of CPU
Manage AI components — Download, update, or delete the AI model

Resource usage

  • Memory: The model increases memory usage while loaded. The English model uses less memory than the Multi Language model.
  • Disk: 32 MB (English) or 117 MB (Multi Language) in %AppData%\Blast\AI\Embeddings\
  • CPU/GPU: Brief spikes during inference. In cached-only mode (no AI tag active), there is minimal overhead.
  • Startup: The model loads asynchronously on startup. A brief warm-up indexes common search tags.

Tips

  • Start with cached mode — leave the AI tag off for everyday use. You'll still benefit from AI-boosted ranking on previously seen results.
  • Use the AI tag when searching by concept — for example, searching for "presentation software" to find PowerPoint.
  • English model is smaller and faster. Use Multi Language only if you search in multiple languages.
  • GPU is worth trying if you have a dedicated graphics card, but CPU works well for the small models Fluent Search uses.