feat(pilot): add binary pruning and pre-filtering for wide nodes

Add binary pruning functionality to quickly filter relevant candidates
before full scoring. Introduce pre-filtering using NodeScorer to reduce
LLM token costs when nodes have many children.

- Add Prune InterventionPoint for binary relevance filtering
- Implement PrefilterConfig and PruneConfig with configurable thresholds
- Add pre-filtering logic in score_candidates_detailed to narrow
  candidate sets using NodeScorer before LLM calls
- Implement binary_prune method in LLM pilot for quick yes/no relevance
  decisions
- Update metrics collection to track pruning interventions
- Add comprehensive tests for new configuration options

Author: zTgx

rust/src/metrics/pilot.rs

@@ -18,6 +18,8 @@ pub enum InterventionPoint {
     Backtrack,
     /// Evaluating content sufficiency.
     Evaluate,
+    /// Binary pruning for wide nodes.
+    Prune,
 }
 
 /// Helper to store f64 as u64 bits for atomic operations.
@@ -87,7 +89,7 @@ impl PilotMetrics {
             InterventionPoint::Start => {
                 self.start_guidance_calls.fetch_add(1, Ordering::Relaxed);
             }
-            InterventionPoint::Fork => {
+            InterventionPoint::Fork | InterventionPoint::Prune => {
                 self.fork_decisions.fetch_add(1, Ordering::Relaxed);
             }
             InterventionPoint::Backtrack => {

rust/src/retrieval/pilot/config.rs

@@ -27,6 +27,10 @@ pub struct PilotConfig {
     pub guide_at_backtrack: bool,
     /// Optional path to custom prompt templates.
     pub prompt_template_path: Option<String>,
+    /// Pre-filtering configuration for reducing candidates before Pilot.
+    pub prefilter: PrefilterConfig,
+    /// Binary pruning configuration for quick relevance filtering.
+    pub prune: PruneConfig,
 }
 
 impl Default for PilotConfig {
@@ -38,6 +42,8 @@ impl Default for PilotConfig {
             guide_at_start: true,
             guide_at_backtrack: true,
             prompt_template_path: None,
+            prefilter: PrefilterConfig::default(),
+            prune: PruneConfig::default(),
         }
     }
 }
@@ -51,7 +57,7 @@ impl PilotConfig {
         }
     }
 
-    /// Create a high-quality config (more LLM calls).
+    /// Create a high-quality config (more LLM calls, generous pre-filter).
     pub fn high_quality() -> Self {
         Self {
             mode: PilotMode::Aggressive,
@@ -71,10 +77,20 @@ impl PilotConfig {
             guide_at_start: true,
             guide_at_backtrack: true,
             prompt_template_path: None,
+            prefilter: PrefilterConfig {
+                threshold: 20,
+                max_to_pilot: 20,
+                enabled: true,
+            },
+            prune: PruneConfig {
+                enabled: true,
+                threshold: 25,
+                min_keep: 5,
+            },
         }
     }
 
-    /// Create a low-cost config (fewer LLM calls).
+    /// Create a low-cost config (fewer LLM calls, aggressive pre-filter).
     pub fn low_cost() -> Self {
         Self {
             mode: PilotMode::Conservative,
@@ -94,13 +110,33 @@ impl PilotConfig {
             guide_at_start: false,
             guide_at_backtrack: true,
             prompt_template_path: None,
+            prefilter: PrefilterConfig {
+                threshold: 8,
+                max_to_pilot: 8,
+                enabled: true,
+            },
+            prune: PruneConfig {
+                enabled: true,
+                threshold: 12,
+                min_keep: 2,
+            },
         }
     }
 
     /// Create a pure algorithm config (no LLM calls).
     pub fn algorithm_only() -> Self {
         Self {
             mode: PilotMode::AlgorithmOnly,
+            prefilter: PrefilterConfig {
+                threshold: 15,
+                max_to_pilot: 15,
+                enabled: false,
+            },
+            prune: PruneConfig {
+                enabled: false,
+                threshold: 20,
+                min_keep: 3,
+            },
             ..Default::default()
         }
     }
@@ -228,6 +264,88 @@ impl InterventionConfig {
     }
 }
 
+/// Configuration for NodeScorer-based pre-filtering before Pilot scoring.
+///
+/// When a node has many children, sending all to the LLM is wasteful.
+/// Pre-filtering uses cheap NodeScorer (keyword/BM25) to narrow the
+/// candidate set before expensive Pilot (LLM) scoring.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct PrefilterConfig {
+    /// Minimum number of candidates to trigger pre-filtering.
+    ///
+    /// When `candidates.len()` exceeds this threshold, NodeScorer
+    /// pre-filters before sending to Pilot.
+    /// Default: 15.
+    pub threshold: usize,
+
+    /// Maximum number of candidates passed to Pilot after pre-filtering.
+    ///
+    /// NodeScorer's top-N are kept; the rest get NodeScorer-only scores.
+    /// Default: 15.
+    pub max_to_pilot: usize,
+
+    /// Whether pre-filtering is enabled.
+    /// Default: true.
+    pub enabled: bool,
+}
+
+impl Default for PrefilterConfig {
+    fn default() -> Self {
+        Self {
+            threshold: 15,
+            max_to_pilot: 15,
+            enabled: true,
+        }
+    }
+}
+
+impl PrefilterConfig {
+    /// Check if pre-filtering should be applied given the candidate count.
+    pub fn should_prefilter(&self, candidate_count: usize) -> bool {
+        self.enabled && candidate_count > self.threshold
+    }
+}
+
+/// Configuration for binary pruning before full Pilot scoring.
+///
+/// After P2 pre-filtering, if candidates still exceed this threshold,
+/// a lightweight LLM call asks "which are relevant?" before the full
+/// scoring call. This reduces the number of candidates that receive
+/// expensive detailed scoring.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct PruneConfig {
+    /// Whether binary pruning is enabled.
+    /// Default: true.
+    pub enabled: bool,
+
+    /// Trigger threshold — binary prune activates when the candidate
+    /// count (after P2 pre-filtering) exceeds this value.
+    /// Default: 20.
+    pub threshold: usize,
+
+    /// Minimum candidates to keep after pruning, even if LLM says
+    /// fewer are relevant. Prevents over-aggressive pruning.
+    /// Default: 3.
+    pub min_keep: usize,
+}
+
+impl Default for PruneConfig {
+    fn default() -> Self {
+        Self {
+            enabled: true,
+            threshold: 20,
+            min_keep: 3,
+        }
+    }
+}
+
+impl PruneConfig {
+    /// Check if binary pruning should be applied given the candidate count.
+    pub fn should_prune(&self, candidate_count: usize) -> bool {
+        self.enabled && candidate_count > self.threshold
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -269,11 +387,68 @@ mod tests {
     fn test_pilot_config_presets() {
         let high = PilotConfig::high_quality();
         assert_eq!(high.mode, PilotMode::Aggressive);
+        assert!(high.prefilter.enabled);
+        assert_eq!(high.prefilter.threshold, 20);
 
         let low = PilotConfig::low_cost();
         assert_eq!(low.mode, PilotMode::Conservative);
+        assert!(low.prefilter.enabled);
+        assert_eq!(low.prefilter.threshold, 8);
 
         let algo = PilotConfig::algorithm_only();
         assert_eq!(algo.mode, PilotMode::AlgorithmOnly);
+        assert!(!algo.prefilter.enabled);
+    }
+
+    #[test]
+    fn test_prefilter_config_default() {
+        let cfg = PrefilterConfig::default();
+        assert!(cfg.enabled);
+        assert_eq!(cfg.threshold, 15);
+        assert_eq!(cfg.max_to_pilot, 15);
+    }
+
+    #[test]
+    fn test_prefilter_should_prefilter() {
+        let cfg = PrefilterConfig::default();
+        assert!(!cfg.should_prefilter(15)); // at threshold
+        assert!(!cfg.should_prefilter(10)); // below
+        assert!(cfg.should_prefilter(16));  // above
+
+        let disabled = PrefilterConfig { enabled: false, ..Default::default() };
+        assert!(!disabled.should_prefilter(100));
+    }
+
+    #[test]
+    fn test_prune_config_default() {
+        let cfg = PruneConfig::default();
+        assert!(cfg.enabled);
+        assert_eq!(cfg.threshold, 20);
+        assert_eq!(cfg.min_keep, 3);
+    }
+
+    #[test]
+    fn test_prune_should_prune() {
+        let cfg = PruneConfig::default();
+        assert!(!cfg.should_prune(20)); // at threshold
+        assert!(!cfg.should_prune(15)); // below
+        assert!(cfg.should_prune(21));  // above
+
+        let disabled = PruneConfig { enabled: false, ..Default::default() };
+        assert!(!disabled.should_prune(100));
+    }
+
+    #[test]
+    fn test_pilot_config_presets_prune() {
+        let high = PilotConfig::high_quality();
+        assert!(high.prune.enabled);
+        assert_eq!(high.prune.threshold, 25);
+
+        let low = PilotConfig::low_cost();
+        assert!(low.prune.enabled);
+        assert_eq!(low.prune.threshold, 12);
+
+        let algo = PilotConfig::algorithm_only();
+        assert!(!algo.prune.enabled);
     }
 }

rust/src/retrieval/pilot/decision.rs

@@ -220,6 +220,8 @@ pub enum InterventionPoint {
     Backtrack,
     /// Evaluating a specific node for relevance.
     Evaluate,
+    /// Binary pruning — quick yes/no relevance filter for wide nodes.
+    Prune,
 }
 
 impl InterventionPoint {
@@ -230,6 +232,7 @@ impl InterventionPoint {
             Self::Fork => "fork",
             Self::Backtrack => "backtrack",
             Self::Evaluate => "evaluate",
+            Self::Prune => "prune",
         }
     }
 }

rust/src/retrieval/pilot/decision_scorer.rs

@@ -112,6 +112,14 @@ pub struct ScoredCandidate {
 /// from the Pilot. Use this when the search algorithm needs to
 /// record why each path step was taken (e.g., for beam search
 /// reasoning history).
+///
+/// # Pre-filtering
+///
+/// When a node has many children (exceeding `prefilter.threshold`),
+/// NodeScorer pre-filters candidates before sending to Pilot. This
+/// reduces LLM token cost and latency. Candidates filtered out still
+/// receive NodeScorer-only scores in the final merge, so no results
+/// are lost.
 pub async fn score_candidates_detailed(
     tree: &DocumentTree,
     candidates: &[NodeId],
@@ -139,20 +147,88 @@ pub async fn score_candidates_detailed(
     // Determine parent node (last in path) for cache key
     let parent = path.last().copied().unwrap_or(tree.root());
 
+    // === PRE-FILTERING ===
+    // When candidates exceed the threshold, use NodeScorer to narrow
+    // the set before sending to Pilot (LLM). Filtered-out candidates
+    // still get NodeScorer-only scores in the final merge below.
+    let prefilter_cfg = &p.config().prefilter;
+    let pilot_candidates: Vec<NodeId> = if prefilter_cfg.should_prefilter(candidates.len()) {
+        let scorer = NodeScorer::new(ScoringContext::new(query));
+        let mut sorted = scorer.score_and_sort(tree, candidates);
+        let pilot_max = prefilter_cfg.max_to_pilot.min(candidates.len());
+        sorted.truncate(pilot_max);
+        let ids: Vec<NodeId> = sorted.into_iter().map(|(id, _)| id).collect();
+        tracing::debug!(
+            "Pre-filtered: {} candidates -> {} to Pilot (threshold={})",
+            candidates.len(),
+            ids.len(),
+            prefilter_cfg.threshold,
+        );
+        ids
+    } else {
+        candidates.to_vec()
+    };
+
+    // === BINARY PRUNING ===
+    // After P2 pre-filtering, if candidates still exceed the prune
+    // threshold, ask Pilot for a quick yes/no filter before the
+    // expensive full-scoring call.
+    let prune_cfg = &p.config().prune;
+    let pilot_candidates = if prune_cfg.should_prune(pilot_candidates.len()) {
+        let mut prune_state = SearchState::new(
+            tree, query, path, &pilot_candidates, visited,
+        );
+        prune_state.step_reasons = step_reasons;
+
+        if let Some(relevant_ids) = p.binary_prune(&prune_state).await {
+            let relevant_set: HashSet<NodeId> = relevant_ids.iter().copied().collect();
+            let mut pruned: Vec<NodeId> = pilot_candidates
+                .iter()
+                .filter(|id| relevant_set.contains(id))
+                .copied()
+                .collect();
+
+            // Enforce min_keep to prevent over-aggressive pruning
+            if pruned.len() < prune_cfg.min_keep {
+                // Fill from the top of pilot_candidates that weren't pruned
+                for id in &pilot_candidates {
+                    if pruned.len() >= prune_cfg.min_keep {
+                        break;
+                    }
+                    if !relevant_set.contains(id) {
+                        pruned.push(*id);
+                    }
+                }
+            }
+
+            tracing::debug!(
+                "Binary prune: {} candidates -> {} relevant (min_keep={})",
+                pilot_candidates.len(),
+                pruned.len(),
+                prune_cfg.min_keep,
+            );
+            pruned
+        } else {
+            pilot_candidates
+        }
+    } else {
+        pilot_candidates
+    };
+
     // Check cache first
     let decision = if let Some(c) = cache {
         if let Some(cached) = c.get(query, parent).await {
             tracing::trace!("Pilot cache hit for parent={:?}", parent);
             cached
         } else {
-            let mut state = SearchState::new(tree, query, path, candidates, visited);
+            let mut state = SearchState::new(tree, query, path, &pilot_candidates, visited);
             state.step_reasons = step_reasons;
             let d = p.decide(&state).await;
             c.put(query, parent, &d).await;
             d
         }
     } else {
-        let mut state = SearchState::new(tree, query, path, candidates, visited);
+        let mut state = SearchState::new(tree, query, path, &pilot_candidates, visited);
         state.step_reasons = step_reasons;
         p.decide(&state).await
     };
@@ -163,7 +239,7 @@ pub async fn score_candidates_detailed(
         pilot_data.insert(ranked.node_id, (ranked.score, ranked.reason.clone()));
     }
 
-    // Compute NodeScorer fallback scores
+    // Compute NodeScorer fallback scores for ALL original candidates
     let scorer_weight = 1.0 - pilot_weight;
     let confidence = decision.confidence;
     let effective_pilot = pilot_weight * confidence;