Skip to content

focus-modes.md

Latest commit

 

History

History
67 lines (52 loc) · 3 KB

focus-modes.md

File metadata and controls

67 lines (52 loc) · 3 KB
Raw

Focus modes

Focus modes tune the discovery and search-aggregation layers toward a specific domain. They bias which connectors run, which engines those connectors hit, and how results get re-ranked before fusion. Where presets shape the output, focus modes shape the input.

Available modes

Mode Source bias Engine selection Recency weight
general (default) Mixed All engines, balanced Moderate
academic .edu, .gov, peer-reviewed, arXiv, SSRN Google Scholar, arxiv, ssrn Low — older sources are fine
documentation Official docs, reference sites Google with site:docs.* boost, Bing Moderate
comparison Comparison sites, blog post round-ups All engines, balanced Moderate
debugging Stack Overflow, GitHub issues, forum posts Google, DuckDuckGo High — recent posts are more likely current
tutorial Blog posts, video transcripts, official guides All engines, balanced Moderate
news News outlets, press releases Google News, Bing News Very high — date-bounded

When to use which

Question type?
├── "What is X?" / "How do I use Y?"
│     → documentation OR tutorial
│
├── "X vs Y" / "Best of"
│     → comparison
│
├── "Why is X erroring with Z?"
│     → debugging
│
├── "Latest research on X"
│     → academic
│
├── "What happened with X this week"
│     → news
│
└── (default / unsure)
      → general

How it works under the hood

src/discovery/focus_modes.py holds the per-mode configuration. A focus mode is a dataclass with:

  • name
  • connector_weights (dict of connector → weight)
  • searxng_engines (override default engines)
  • recency_weight (multiplier on RRF fusion for recent results)
  • domain_boosts (list of (domain_pattern, boost_factor))
  • domain_penalties (list of (domain_pattern, penalty_factor))

When a request specifies a focus mode, the discovery layer:

  1. Picks the connectors with non-zero weight.
  2. Overrides each connector's engine selection.
  3. Adjusts the post-fusion ranking using recency_weight × domain_boosts/penalties.

The synthesis layer also reads the focus mode to adjust prompt templates — e.g. academic uses a more formal system prompt with explicit citation-format instructions.

Combining focus modes and presets

See the preset-mode crosswalk in presets.md. The two axes are orthogonal: pick the preset based on the answer shape you want (fast vs comprehensive vs comparison), and the focus mode based on where the answer lives (academic vs forum vs official docs).

Adding a new focus mode

  1. Add a new entry to src/discovery/focus_modes.py following the dataclass pattern.
  2. Add a corresponding system-prompt template under src/synthesis/prompts/focus_modes/<name>.md if the synthesis stage should behave differently for this mode.
  3. The mode is auto-exposed at /api/v1/focus-modes and via the focus_mode argument to discover, synthesize, and reason.