forgeplan fpf search

forgeplan fpf search queries the ingested First Principles Framework knowledge base using the same hybrid retrieval pipeline that powers artifact search — production BM25 (via the bm25 crate v2.3.2) fused with BGE-M3 semantic vectors. It searches across all 204 FPF sections and returns ranked matches with section IDs and previews.

Think of it as forgeplan search pointed at the FPF corpus instead of your own artifacts.

When to use

While shaping a PRD or ADR — pull first-principles context on a concept before writing ("trust calculus", "bounded context", "explore exploit").
While auditing a decision — verify the framework’s stance on a pattern you’re considering.
While onboarding — treat the KB as a searchable textbook.
While writing forgeplan reason --fpf prompts — preview what the model will see as FPF context.

When NOT to use

For artifact search — use forgeplan search instead.
For reading a whole section — use forgeplan fpf section <id>.
For browsing the full index — use forgeplan fpf list.

Usage

forgeplan fpf search [OPTIONS] <QUERY>

Arguments

  <QUERY>  Search query

Options

      --limit <LIMIT>  Max results [default: 5]
      --semantic       Use semantic vector search (requires --features semantic-search;
                       falls back to keyword otherwise)
  -h, --help           Print help
  -V, --version        Print version

Examples

# Default: BM25 keyword search, top 5 results
forgeplan fpf search "trust calculus"

# Widen the result window
forgeplan fpf search "explore exploit tradeoff" --limit 10

# Semantic vector search (falls back to BM25 if semantic-search feature is off)
forgeplan fpf search "adversarial review" --semantic

forgeplan fpf search "weakest link"

How the pipeline works

The FPF search uses the same v0.18.0 production pipeline as artifact search:

BM25 (bm25 crate v2.3.2) — tokenization with Russian morphology (Snowball stemmer) when LanguageMode::Detect triggers.
Template noise stripping — removes boilerplate headings before scoring, so queries match substantive content.
Semantic fusion — BGE-M3 cosine similarity layered on top when the semantic-search feature is enabled.
O(N) batch search — efficient across the 204-section corpus.

Because the mechanics are identical, the same tuning and troubleshooting notes from the search v2 guide apply here.

How it fits

fpf search is the query surface of the FPF KB. It’s consumed both by humans (CLI inspection) and by forgeplan reason --fpf, which retrieves top-k sections as context for the ADI prompt.

ingest → search → (human reads OR reason consumes) → better decisions