Web Search

FlowDown offers two web search modes: pre-processing for models without tool-calling, and tool-calls where the assistant invokes web_search directly. When your model understands tools, turn on both Web Browsing and Tools so it can plan and cite on its own.

Enable Web Search

  • Go to Settings → Tools → Web Search or toggle Web Browsing above the composer.
  • For tool-capable models: turn on both Web Browsing and Tools. FlowDown skips pre-planning and lets the model call web_search.
  • For models without tool-calling: leave Tools off and enable Web Browsing. FlowDown asks the auxiliary model to decide and draft queries.

Configuration Guide

  • Search Strategy (Essential / Balanced / Proactive): Essential searches only when needed; Balanced balances frequency and speed; Proactive searches aggressively for more coverage. Higher sensitivity can add latency, and lower sensitivity may miss fresh info.

Web search strategy and engine configuration

  • Search Limit (5/10/15/20/Unlimited pages): shared across all queries in one message. Pages beyond the context budget are dropped to prevent overflow.

Web search parameter configuration interface

  • Web Search Engines lets you toggle Google / DuckDuckGo / Yahoo / Bing. Keep at least one on; if all are off, FlowDown automatically falls back to Google.

Execution Flow

Pre-processing mode (Tools off)

  1. Context collection – Last five user/assistant turns plus attachment text.
  2. Query planning – Auxiliary model applies the Search Strategy prompt and returns XML with search_required and up to three queries (2–25 characters, in the user’s language). If search_required is false, FlowDown posts a “no web search needed” note.
  3. Failure handling – If no queries are produced, FlowDown replies that it cannot generate search keywords and skips browsing.
  4. Retrieval – ScrubberKit + URLSReranker fetch per query (keeps four results per host) and honor the per-message limit, streaming sources/sites/result counts.
  5. Attachment injection – Results are shuffled, converted to numbered web_document attachments, and appended before the model answers. Empty results raise “No web search results.”

Tool-call mode (Tools on)

  • When the model calls web_search, FlowDown runs the same ScrubberKit pipeline and streams progress in the status panel. The tool response includes web_document text; if nothing is found, it returns “Web search returned no results.”

Use Cases

  • Track fast-moving news or release notes.
  • Look up pricing, availability, or policies that change frequently.
  • Enrich domain knowledge with recent publications or documentation.
  • Capture structured data (tables, lists) from public pages.

Best Practices

Ask Effectively

  • Provide context, e.g., “Find recent reviews for the latest FlowDown release.”
  • Add time ranges or preferred sources, e.g., “Look for docs updated after October 2025.”
  • Combine with attachments or notes and ask the assistant to cite web_document IDs with [^n].

Search Tuning

  • Use Balanced or Proactive for dynamic topics; choose Essential for speed.
  • Keep at least one engine enabled; more engines improve recall but add latency.
  • Raise Search Limit for sparse or long pages; lower it to conserve bandwidth.
  • Review URLs and snippets in the reasoning panel before trusting an answer.

Privacy Notes

  • FlowDown only sends the minimal query terms to providers and does not append personal identifiers.
  • Scrubbed content stays in your local conversation history and syncs via iCloud only if enabled.
  • Disable web search entirely if your compliance requirements forbid contacting external providers.

Implementation Details

  • Query generation uses the auxiliary model with the Search Strategy prompt (Model.Inference.SearchSensitivity).
  • Retrieval uses ScrubberKit plus URLSReranker and respects engine toggles and search-limit settings (ScrubberConfiguration).
  • Tool-call entry point: MTWebSearchTool runs the pipeline, and the conversation layer wraps results with formatAsWebArchive so the model can cite web_document IDs.