glpi-neural-brain/internal/graph/vector_layer.go

package graph

import (
	"fmt"
	"math"
	"sort"
	"time"

	"github.com/local/glpi-neural-brain/internal/model"
	"github.com/local/glpi-neural-brain/internal/vectorgraph"
)

const VectorMathOrigin = "vector-math"

type VectorSemanticLayerConfig struct {
	Workers        int
	Neighbors      int
	CandidateLimit int
	HashBits       int
	HashTables     int
	MinSimilarity  float64
	MinAffinity    float64
	Layout         bool
	LayoutRelax    bool
	LayoutBlend    float64
	LayoutMaxShift float64

	OrphanPass           bool
	OrphanNeighbors      int
	OrphanCandidateLimit int
	OrphanMinSimilarity  float64
	OrphanMinAffinity    float64
}

type VectorSemanticLayerStats struct {
	vectorgraph.Stats
	OrphanStats     vectorgraph.Stats `json:"orphan_stats"`
	OrphanFocus     int               `json:"orphan_focus"`
	OrphanLinks     int               `json:"orphan_links"`
	PositionUpdates uint64            `json:"position_updates"`
}

func (s *Store) HasEdgesByOrigin(origin string) bool {
	s.mu.RLock()
	defer s.mu.RUnlock()
	for _, edge := range s.edges {
		if edge.Origin == origin && edge.Status != "rejected" {
			return true
		}
	}
	return false
}

// VectorSemanticEntries snapshots the production knowledge vectors used by the
// deterministic semantic layer. Callers may safely ship this immutable copy to
// a CPU worker/Agent while the live graph stays owned by the Brain process.
func (s *Store) VectorSemanticEntries(filter NodeFilter) []vectorgraph.Entry {
	s.mu.RLock()
	defer s.mu.RUnlock()
	entries := make([]vectorgraph.Entry, 0, len(s.vectors))
	for id, node := range s.nodes {
		if node.Kind != "knowledge" || node.Status != "production" || !filter.Matches(node) {
			continue
		}
		vector, ok := s.vectors[id]
		if !ok || len(vector) == 0 {
			continue
		}
		entries = append(entries, vectorgraph.Entry{ID: id, Vector: append([]float32(nil), vector...)})
	}
	sort.Slice(entries, func(i, j int) bool { return entries[i].ID < entries[j].ID })
	return entries
}

// KnowledgeOrphanIDsIgnoringOrigin returns production knowledge nodes that have
// no direct Knowledge/External evidence relationship when edges from the given
// origin are ignored. The vector-math layer uses this before rebuilding itself
// so an old mathematical edge does not hide a genuine orphan from pass two.
func (s *Store) KnowledgeOrphanIDsIgnoringOrigin(filter NodeFilter, ignoredOrigin string) []string {
	s.mu.RLock()
	defer s.mu.RUnlock()
	linked := map[string]bool{}
	for _, edge := range s.edges {
		if edge.Status == "rejected" || edge.Origin == ignoredOrigin {
			continue
		}
		a, aok := s.nodes[edge.Source]
		b, bok := s.nodes[edge.Target]
		if !aok || !bok {
			continue
		}
		aKnowledge := a.Kind == "knowledge" && a.Status == "production" && filter.Matches(a)
		bKnowledge := b.Kind == "knowledge" && b.Status == "production" && filter.Matches(b)
		if aKnowledge && (b.Kind == "knowledge" || b.Kind == "ai-think" || b.Kind == "external") {
			linked[a.ID] = true
		}
		if bKnowledge && (a.Kind == "knowledge" || a.Kind == "ai-think" || a.Kind == "external") {
			linked[b.ID] = true
		}
	}
	out := make([]string, 0)
	for id, node := range s.nodes {
		if node.Kind == "knowledge" && node.Status == "production" && filter.Matches(node) && !linked[id] {
			if vector := s.vectors[id]; len(vector) > 0 {
				out = append(out, id)
			}
		}
	}
	sort.Strings(out)
	return out
}

func vectorBuildConfig(cfg VectorSemanticLayerConfig) vectorgraph.Config {
	return vectorgraph.Config{
		Workers: cfg.Workers, Neighbors: cfg.Neighbors, CandidateLimit: cfg.CandidateLimit,
		HashBits: cfg.HashBits, HashTables: cfg.HashTables, BandBits: 8,
		MinSimilarity: cfg.MinSimilarity, MinAffinity: cfg.MinAffinity,
		Layout: cfg.Layout, Smoothing: .22,
	}
}

func orphanVectorBuildConfig(cfg VectorSemanticLayerConfig) vectorgraph.Config {
	return vectorgraph.Config{
		Workers: cfg.Workers, Neighbors: cfg.OrphanNeighbors, CandidateLimit: cfg.OrphanCandidateLimit,
		HashBits: cfg.HashBits, HashTables: cfg.HashTables, BandBits: 8,
		MinSimilarity: cfg.OrphanMinSimilarity, MinAffinity: cfg.OrphanMinAffinity,
		Layout: false,
	}
}

// BuildVectorSemanticLayerLocal performs both mathematical passes locally. It
// does not mutate the graph and can therefore be replaced transparently by an
// Agent result using the same inputs/configuration.
func (s *Store) BuildVectorSemanticLayerLocal(cfg VectorSemanticLayerConfig, filter NodeFilter) (vectorgraph.Result, vectorgraph.Result, []string) {
	entries := s.VectorSemanticEntries(filter)
	baseOrphans := s.KnowledgeOrphanIDsIgnoringOrigin(filter, VectorMathOrigin)
	primary := vectorgraph.Build(entries, vectorBuildConfig(cfg))
	if !cfg.OrphanPass || len(baseOrphans) == 0 {
		return primary, vectorgraph.Result{}, nil
	}
	focus := make(map[string]bool, len(baseOrphans))
	for _, id := range baseOrphans {
		focus[id] = true
	}
	for _, link := range primary.Links {
		delete(focus, link.Source)
		delete(focus, link.Target)
	}
	focusIDs := make([]string, 0, len(focus))
	for id := range focus {
		focusIDs = append(focusIDs, id)
	}
	sort.Strings(focusIDs)
	orphan := vectorgraph.BuildFocused(entries, focus, orphanVectorBuildConfig(cfg))
	return primary, orphan, focusIDs
}

// ApplyVectorSemanticLayer atomically replaces the mathematical edge layer
// with a result computed either locally or on an Agent. The Brain remains the
// sole graph owner and validates endpoint existence while applying the result.
func (s *Store) ApplyVectorSemanticLayer(cfg VectorSemanticLayerConfig, primary, orphan vectorgraph.Result, orphanFocus []string) (VectorSemanticLayerStats, MutationStats) {
	type taggedLink struct {
		link vectorgraph.Link
		pass string
	}
	byPair := map[string]taggedLink{}
	for _, link := range primary.Links {
		byPair[pairKey(link.Source, link.Target)] = taggedLink{link: link, pass: "primary"}
	}
	for _, link := range orphan.Links {
		key := pairKey(link.Source, link.Target)
		if _, exists := byPair[key]; !exists {
			byPair[key] = taggedLink{link: link, pass: "orphan"}
		}
	}
	keys := make([]string, 0, len(byPair))
	for key := range byPair {
		keys = append(keys, key)
	}
	sort.Strings(keys)
	edges := make([]model.Edge, 0, len(keys))
	now := time.Now().UTC()
	for _, key := range keys {
		tagged := byPair[key]
		link := tagged.link
		algorithm := "mutual-knn-local-scaling-v1"
		if tagged.pass == "orphan" {
			algorithm = "orphan-knn-local-scaling-v1"
		}
		edges = append(edges, model.Edge{
			Source: link.Source, Target: link.Target, Type: "semantic_neighbor", Origin: VectorMathOrigin,
			Status: "staging", Confidence: link.Confidence, Weight: math.Max(.2, link.Affinity),
			Explanation: fmt.Sprintf("Deterministische Vektornachbarschaft (%s): Cosine %.4f, lokal skalierte Affinität %.4f", tagged.pass, link.Similarity, link.Affinity),
			Metadata: map[string]any{
				"algorithm": algorithm, "pass": tagged.pass, "semantic_similarity": link.Similarity,
				"local_affinity": link.Affinity, "reciprocal": link.Reciprocal,
				"source_rank": link.SourceRank, "target_rank": link.TargetRank, "no_model_call": true,
			},
			CreatedAt: now, UpdatedAt: now,
		})
	}
	mutations := s.ReplaceOriginsWithStats([]string{VectorMathOrigin}, nil, edges)
	stats := VectorSemanticLayerStats{Stats: primary.Stats, OrphanStats: orphan.Stats, OrphanFocus: len(orphanFocus), OrphanLinks: len(orphan.Links)}
	if cfg.Layout && len(primary.Positions) > 0 {
		var updated MutationStats
		if cfg.LayoutRelax {
			updated = s.applyVectorPositionsRelaxed(primary.Positions, cfg.LayoutBlend, cfg.LayoutMaxShift)
		} else {
			updated = s.applyVectorPositions(primary.Positions)
		}
		stats.PositionUpdates = updated.NodesUpdated
		mutations.Add(updated)
	}
	return stats, mutations
}

// RebuildVectorSemanticLayer creates a sparse Knowledge<->Knowledge semantic
// layer from already stored embeddings. It performs no model/network request.
// The generated relation is intentionally named semantic_neighbor rather than
// same_topic: vector proximity is a mathematical neighbourhood signal, not a
// factual relation decision.
func (s *Store) RebuildVectorSemanticLayer(cfg VectorSemanticLayerConfig, filter NodeFilter) (VectorSemanticLayerStats, MutationStats) {
	primary, orphan, focus := s.BuildVectorSemanticLayerLocal(cfg, filter)
	return s.ApplyVectorSemanticLayer(cfg, primary, orphan, focus)
}

func (s *Store) applyVectorPositionsRelaxed(positions []vectorgraph.Position, blend, maxShift float64) MutationStats {
	if blend <= 0 {
		blend = .08
	}
	if blend > .5 {
		blend = .5
	}
	if maxShift <= 0 {
		maxShift = .035
	}
	wanted := make(map[string]vectorgraph.Position, len(positions))
	for _, p := range positions {
		wanted[p.ID] = p
	}
	s.mu.Lock()
	defer s.mu.Unlock()
	var stats MutationStats
	for id, p := range wanted {
		node, ok := s.nodes[id]
		if !ok || node.Kind != "knowledge" || node.Status != "production" {
			continue
		}
		tx, ty, tz := fitVectorPosition(node.ID, p.X, p.Y, p.Z)
		dx, dy, dz := (tx-node.X)*blend, (ty-node.Y)*blend, (tz-node.Z)*blend
		distance := math.Sqrt(dx*dx + dy*dy + dz*dz)
		// Do not dirty thousands of rows for sub-pixel relaxation noise. Periodic
		// reevaluation will revisit the target later if the semantic geometry moves.
		if distance < .00075 {
			continue
		}
		if distance > maxShift && distance > 0 {
			scale := maxShift / distance
			dx, dy, dz = dx*scale, dy*scale, dz*scale
		}
		x, y, z := node.X+dx, node.Y+dy, node.Z+dz
		if math.Abs(node.X-x) < 1e-7 && math.Abs(node.Y-y) < 1e-7 && math.Abs(node.Z-z) < 1e-7 {
			continue
		}
		old := node
		node.X, node.Y, node.Z = x, y, z
		s.nodes[id] = node
		s.countNodeUpdatedLocked()
		stats.NodesUpdated++
		s.version++
		s.recordChangeLocked(nodeUpdateChange(old, node))
		s.markNodeDirtyLocked(id)
	}
	return stats
}

func (s *Store) applyVectorPositions(positions []vectorgraph.Position) MutationStats {
	wanted := make(map[string]vectorgraph.Position, len(positions))
	for _, p := range positions {
		wanted[p.ID] = p
	}
	s.mu.Lock()
	defer s.mu.Unlock()
	var stats MutationStats
	for id, p := range wanted {
		node, ok := s.nodes[id]
		if !ok || node.Kind != "knowledge" || node.Status != "production" {
			continue
		}
		x, y, z := fitVectorPosition(node.ID, p.X, p.Y, p.Z)
		if math.Abs(node.X-x) < 1e-7 && math.Abs(node.Y-y) < 1e-7 && math.Abs(node.Z-z) < 1e-7 {
			continue
		}
		old := node
		node.X, node.Y, node.Z = x, y, z
		// Position is a derived visualization property; preserve source freshness.
		s.nodes[id] = node
		s.countNodeUpdatedLocked()
		stats.NodesUpdated++
		s.version++
		s.recordChangeLocked(nodeUpdateChange(old, node))
		s.markNodeDirtyLocked(id)
	}
	return stats
}

func fitVectorPosition(id string, x, y, z float64) (float64, float64, float64) {
	x = math.Max(-.82, math.Min(.82, x))
	y = math.Max(-.78, math.Min(.82, y))
	z = math.Max(-.62, math.Min(.62, z))
	if math.Abs(x) < .055 && y > -.58 && y < .42 {
		if ID("vector-layout-side", id)[0]%2 == 0 {
			x = .06
		} else {
			x = -.06
		}
	}
	for i := 0; i < 20 && !insideBrainShape(x, y, z); i++ {
		x *= .94
		y *= .94
		z *= .94
		if math.Abs(x) < .055 && y > -.58 && y < .42 {
			if x >= 0 {
				x = .06
			} else {
				x = -.06
			}
		}
	}
	if !insideBrainShape(x, y, z) {
		return position(id, nil)
	}
	return x, y, z
}

type VectorNeighborCandidateStats struct {
	Candidates      int `json:"candidates"`
	AlreadyReviewed int `json:"already_reviewed"`
}

// NextVectorNeighborPairScoped returns the strongest mathematical neighbourhood
// that has not yet been reviewed by AI-THINK. This turns the vector layer into
// a cheap candidate generator: the LLM evaluates relations instead of spending
// model time searching the full embedding space again.
func (s *Store) NextVectorNeighborPairScoped(filter NodeFilter, maxAIDepth int) (model.Node, model.Node, float64, bool, VectorNeighborCandidateStats) {
	s.mu.RLock()
	defer s.mu.RUnlock()
	stats := VectorNeighborCandidateStats{}
	reviewed := map[string]bool{}
	for _, edge := range s.edges {
		if edge.Origin == "ai-inference" {
			reviewed[pairKey(edge.Source, edge.Target)] = true
		}
	}
	type candidate struct {
		edge       model.Edge
		similarity float64
		reciprocal bool
	}
	candidates := make([]candidate, 0)
	for _, edge := range s.edges {
		if edge.Origin != VectorMathOrigin || edge.Type != "semantic_neighbor" || edge.Status == "rejected" {
			continue
		}
		a, aok := s.nodes[edge.Source]
		b, bok := s.nodes[edge.Target]
		if !aok || !bok || !filter.Matches(a) || !filter.Matches(b) {
			continue
		}
		if a.Kind != "knowledge" || b.Kind != "knowledge" {
			continue
		}
		if (a.Kind == "ai-think" && maxAIDepth > 0 && graphNodeGenerationDepth(a) >= maxAIDepth) || (b.Kind == "ai-think" && maxAIDepth > 0 && graphNodeGenerationDepth(b) >= maxAIDepth) {
			continue
		}
		stats.Candidates++
		if reviewed[pairKey(a.ID, b.ID)] {
			stats.AlreadyReviewed++
			continue
		}
		similarity := edge.Confidence
		if value, ok := edge.Metadata["semantic_similarity"].(float64); ok {
			similarity = value
		} else if value, ok := edge.Metadata["semantic_similarity"].(float32); ok {
			similarity = float64(value)
		}
		reciprocal, _ := edge.Metadata["reciprocal"].(bool)
		candidates = append(candidates, candidate{edge: edge, similarity: similarity, reciprocal: reciprocal})
	}
	if len(candidates) == 0 {
		return model.Node{}, model.Node{}, 0, false, stats
	}
	sort.Slice(candidates, func(i, j int) bool {
		if candidates[i].reciprocal != candidates[j].reciprocal {
			return candidates[i].reciprocal
		}
		if candidates[i].edge.Confidence != candidates[j].edge.Confidence {
			return candidates[i].edge.Confidence > candidates[j].edge.Confidence
		}
		if candidates[i].similarity != candidates[j].similarity {
			return candidates[i].similarity > candidates[j].similarity
		}
		return candidates[i].edge.ID < candidates[j].edge.ID
	})
	best := candidates[0]
	return s.nodes[best.edge.Source], s.nodes[best.edge.Target], best.similarity, true, stats
}