Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

28.6 KB
Raw Blame
//
//  Copyright 2026 The InfiniFlow Authors. All Rights Reserved.
//
//  Licensed under the Apache License, Version 2.0 (the "License");
//  you may not use this file except in compliance with the License.
//  You may obtain a copy of the License at
//
//      http://www.apache.org/licenses/LICENSE-2.0
//
//  Unless required by applicable law or agreed to in writing, software
//  distributed under the License is distributed on an "AS IS" BASIS,
//  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
//  See the License for the specific language governing permissions and
//  limitations under the License.
//

package agentic_rag

import (
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"sort"
	"strings"
	"sync"

	"ragflow/internal/engine"
	enginetypes "ragflow/internal/engine/types"
)

// Knowledge-graph walk backing the graph_explore tool: seed entity rows for
// the query, hop out over relation rows, and hand the subgraph plus the
// source passages behind it back to the model.
//
// The walk never calls a chat model. In this ReAct loop the model is already
// the judge of its own evidence, a tool holds no chat model, and an
// answer-only result would strip the passages the answer has to cite. The
// walk therefore always returns the subgraph and its passages.

// Error returned when no KgService was installed at boot.
var ErrKgServiceMissing = errors.New(
	"kg service not registered — call agentic_rag.SetKgService(...) at boot",
)

// Kg row scopes and walk limits, mirroring the harness constants.
const (
	kgScopeDataset = "dataset"
	kgScopeDoc     = "doc"

	kgSeeds       = 2   // top-N entities matched directly to the question
	kgSeedPool    = 64  // KNN candidate pool before the mention_count re-sort
	kgSeedSim     = 0.8 // dense seed similarity floor
	kgHops        = 2   // relation hops out from the seeds
	kgNeighbors   = 128 // cap on neighbour entity rows resolved per hop
	kgRelLimit    = 32  // relations fetched per endpoint filter
	kgMaxPassages = 6   // source passages returned per walk

	kgOutlineMaxDocs  = 3 // documents per outline call
	kgOutlineRowsEach = 1 // a document has exactly one `graph` row
	// kgOutlineMaxEntities / kgOutlineMaxRelations bound the skeleton to what
	// the model can actually use: the hubs. An outline enters the ReAct
	// transcript and is re-sent on EVERY later turn, so a 60-entity dump costs
	// more than the deep read it was meant to replace. The mention-count sort
	// already puts the useful entities first, so cutting the tail is free.
	kgOutlineMaxEntities  = 15
	kgOutlineMaxRelations = 20

	kgPassageRunes = 1500
)

// KgEntity is one entity row of the compiled graph.
type KgEntity struct {
	Name           string
	Type           string
	Description    string
	Aliases        []string
	SourceChunkIDs []string
	DocID          string
	// MentionCount is how often the corpus mentions the entity: a centrality
	// proxy that lets the caller tell a hub from a leaf.
	MentionCount int
}

// KgRelation is one relation (edge) row of the compiled graph.
type KgRelation struct {
	ID             string
	From           string
	To             string
	Type           string
	SourceChunkIDs []string
	DocID          string
}

// KgPassage is one source chunk behind a graph node or edge.
type KgPassage struct {
	ChunkID   string
	DocID     string
	DocName   string
	DatasetID string
	Content   string
}

// KgExploreRequest is one graph walk request. DatasetIDs is the conversation's
// server-bound scope; DocScope, when non-empty, pins the walk to those docs.
type KgExploreRequest struct {
	TenantID   string
	DatasetIDs []string
	DocScope   []string
	Query      string
}

// KgExploreResult is the walk output: the seeded + hopped subgraph and the
// passages behind it, in first-seen order.
type KgExploreResult struct {
	Entities  []KgEntity
	Relations []KgRelation
	Passages  []KgPassage
	// GraphAbsent is true ONLY when the bound scope provably holds no compiled
	// entity rows at all - a dataset-level fact that does not change within a
	// conversation, which is what lets the tool disable itself. An empty walk
	// with GraphAbsent false is a query-level miss (the graph exists but this
	// query seeded nothing) or an infrastructure failure: neither may disable
	// the tool, because a better-phrased query may still hit.
	GraphAbsent bool
}

// KgService walks the compiled knowledge graph. The tools resolve it per call
// so the server can install the engine-backed adapter at boot.
type KgService interface {
	Explore(ctx context.Context, req KgExploreRequest) (KgExploreResult, error)
	DocOutline(ctx context.Context, req KgDocOutlineRequest) (KgExploreResult, error)
}

var (
	kgServiceMu   sync.RWMutex
	kgServiceImpl KgService = stubKgService{}
)

// SetKgService installs the graph-walk service. nil restores the stub.
func SetKgService(svc KgService) {
	kgServiceMu.Lock()
	defer kgServiceMu.Unlock()
	if svc == nil {
		kgServiceImpl = stubKgService{}
		return
	}
	kgServiceImpl = svc
}

// GetKgService returns the installed graph-walk service (a failing stub when
// none was installed).
func GetKgService() KgService {
	kgServiceMu.RLock()
	defer kgServiceMu.RUnlock()
	return kgServiceImpl
}

type stubKgService struct{}

func (stubKgService) Explore(context.Context, KgExploreRequest) (KgExploreResult, error) {
	return KgExploreResult{}, ErrKgServiceMissing
}

func (stubKgService) DocOutline(context.Context, KgDocOutlineRequest) (KgExploreResult, error) {
	return KgExploreResult{}, ErrKgServiceMissing
}

// KgAdapter walks the compiled graph rows through the document engine. It is
// stateless and safe to share across goroutines.
type KgAdapter struct {
	docEngine engine.DocEngine
	// seedEncoder embeds the query for the dense seed search. nil (or a nil
	// vector) degrades the seed search to a keyword match on the entity rows,
	// so the walk still works on a deployment without a tenant embedder.
	seedEncoder func(ctx context.Context, tenantID, text string) []float64
}

// NewKgAdapter wraps a doc engine behind the KgService interface.
func NewKgAdapter(docEngine engine.DocEngine) *KgAdapter {
	return &KgAdapter{docEngine: docEngine}
}

// SetSeedEncoder installs the query-side embedder used by the dense seed
// search. Pass nil to fall back to keyword seeding.
func (a *KgAdapter) SetSeedEncoder(fn func(ctx context.Context, tenantID, text string) []float64) {
	if a == nil {
		return
	}
	a.seedEncoder = fn
}

// Explore seeds entities for the query, hops kgHops out over their relations,
// and returns the subgraph with the passages behind it.
func (a *KgAdapter) Explore(ctx context.Context, req KgExploreRequest) (KgExploreResult, error) {
	empty := KgExploreResult{}
	if a == nil || a.docEngine == nil {
		return empty, ErrKgServiceMissing
	}
	query := strings.TrimSpace(req.Query)
	datasetIDs := nonEmptyStrings(req.DatasetIDs)
	if query == "" || len(datasetIDs) == 0 {
		return empty, nil
	}

	// Encode the seed text ONCE: one embedding request serves every dataset.
	seedVec := a.seedVector(ctx, req.TenantID, query)

	var entities []KgEntity
	var relations []KgRelation
	seenNames := map[string]bool{}

	addEntities := func(found []KgEntity, kbID string) []string {
		var added []string
		for _, e := range found {
			key := kbID + ":" + strings.ToLower(e.Name)
			if seenNames[key] {
				continue
			}
			seenNames[key] = true
			entities = append(entities, e)
			added = append(added, e.Name)
		}
		return added
	}

	docs := nonEmptyStrings(req.DocScope)
	// Whole-dataset walk: no scope filter. The compacted rows of this corpus
	// live at scope_kwd=doc, so filtering by scope_kwd=dataset would read the
	// graph as empty; the harness's dataset default is not carried over.
	scopeKwd := ""
	if len(docs) > 0 {
		scopeKwd = kgScopeDoc
	}

	for _, kbID := range datasetIDs {
		seedRows := a.seedSearch(ctx, req.TenantID, kbID, docs, query, scopeKwd, seedVec)
		var seeds []KgEntity
		for _, row := range seedRows {
			if e, ok := kgParseEntity(row); ok {
				seeds = append(seeds, e)
			}
		}
		frontier := addEntities(seeds, kbID)

		for hop := 0; hop < kgHops; hop++ {
			if len(frontier) == 0 {
				break
			}
			terms := kgEndpointTerms(frontier)
			relRows := a.kgSearch(ctx, req.TenantID, kbID, docs, "relation", "", kgRelLimit, scopeKwd,
				map[string]interface{}{"from_entity_kwd": terms}, "", 0)
			relRows = append(relRows, a.kgSearch(ctx, req.TenantID, kbID, docs, "relation", "", kgRelLimit, scopeKwd,
				map[string]interface{}{"to_entity_kwd": terms}, "", 0)...)

			// Dedup on the row id so the same endpoint pair from different rows
			// survives; fall back to the from|to|type triple when the engine
			// returned no id.
			seenRel := map[string]bool{}
			var hopRelations []KgRelation
			for _, row := range relRows {
				rel, ok := kgParseRelation(row)
				if !ok {
					continue
				}
				key := rel.ID
				if key == "" {
					key = rel.From + "|" + rel.To + "|" + rel.Type
				}
				if seenRel[key] {
					continue
				}
				seenRel[key] = true
				hopRelations = append(hopRelations, rel)
			}
			relations = append(relations, hopRelations...)

			seenLower := map[string]bool{}
			for key := range seenNames {
				if strings.HasPrefix(key, kbID+":") {
					seenLower[strings.TrimPrefix(key, kbID+":")] = true
				}
			}
			neighSet := map[string]string{}
			for _, rel := range hopRelations {
				for _, name := range []string{rel.From, rel.To} {
					name = strings.TrimSpace(name)
					if name == "" || seenLower[strings.ToLower(name)] {
						continue
					}
					neighSet[strings.ToLower(name)] = name
				}
			}
			if len(neighSet) == 0 {
				break
			}
			neighbours := make([]string, 0, len(neighSet))
			for _, name := range neighSet {
				neighbours = append(neighbours, name)
			}
			limit := kgNeighbors
			if len(neighbours) < limit {
				limit = len(neighbours)
			}
			neighRows := a.kgSearch(ctx, req.TenantID, kbID, docs, "entity", "", limit, scopeKwd,
				map[string]interface{}{"name_kwd": kgEndpointTerms(neighbours)}, "", 0)
			var found []KgEntity
			for _, row := range neighRows {
				if e, ok := kgParseEntity(row); ok {
					found = append(found, e)
				}
			}
			frontier = addEntities(found, kbID)
		}
	}

	if len(entities) == 0 && len(relations) == 0 {
		// A doc-scoped walk (the model passed doc_scope) can only ever prove
		// that THOSE documents hold no entity row - it says nothing about the
		// dataset, so it must not report a dataset-level absence: that would
		// disable graph_explore for the whole conversation on a single scoped
		// miss. Report a plain empty result instead, which the tool treats as a
		// query miss.
		if len(docs) > 0 {
			return empty, nil
		}
		// An unscoped empty walk is still ambiguous: the dataset may hold a
		// graph that this query seeded nothing from, or no compiled graph at
		// all. Probe the entity rows to tell them apart - and propagate a probe
		// failure rather than misreading an infrastructure error as "no graph".
		// Only a probe that succeeded and found nothing is a dataset-level
		// absence.
		for _, kbID := range datasetIDs {
			has, probeErr := a.graphRowsExist(ctx, req.TenantID, kbID, nil, "")
			if probeErr != nil {
				return empty, probeErr
			}
			if has {
				return empty, nil
			}
		}
		return KgExploreResult{GraphAbsent: true}, nil
	}
	return KgExploreResult{
		Entities:  entities,
		Relations: relations,
		Passages:  a.evidencePassages(ctx, req.TenantID, datasetIDs, entities, relations),
	}, nil
}

// seedVector encodes the query for the dense seed search. A nil result (no
// encoder, or an encoding failure) sends the caller to the keyword fallback.
func (a *KgAdapter) seedVector(ctx context.Context, tenantID, text string) []float64 {
	if a.seedEncoder == nil {
		return nil
	}
	defer func() {
		// A misconfigured embedding service must degrade to keyword seeding,
		// never take the turn down.
		_ = recover()
	}()
	vec := a.seedEncoder(ctx, tenantID, text)
	if len(vec) == 0 {
		return nil
	}
	return vec
}

// seedSearch resolves the seed entity rows: dense KNN over the entity rows
// re-ranked by mention_count desc, or a keyword match when no vector is
// available.
func (a *KgAdapter) seedSearch(ctx context.Context, tenantID, kbID string, docs []string, text, scopeKwd string, seedVec []float64) []map[string]interface{} {
	if len(seedVec) > 0 {
		dense := &enginetypes.MatchDenseExpr{
			VectorColumnName:  fmt.Sprintf("q_%d_vec", len(seedVec)),
			EmbeddingData:     seedVec,
			EmbeddingDataType: "float",
			DistanceType:      "cosine",
			TopN:              kgSeedPool,
			ExtraOptions:      map[string]interface{}{"similarity": kgSeedSim},
		}
		rows := a.kgSearchRaw(ctx, tenantID, kbID, docs, "entity", scopeKwd, nil, []interface{}{dense}, "mention_count_int", kgSeedPool)
		return kgTopMentionCount(rows, kgSeeds)
	}
	return a.kgSearch(ctx, tenantID, kbID, docs, "entity", text, kgSeeds, scopeKwd, nil, "mention_count_int", kgSeedPool)
}

// graphRowsExist reports whether one dataset holds ANY compiled entity row in
// the given scope. Unlike kgSearchRaw it PROPAGATES the engine error: the caller
// uses the answer to decide whether an empty walk is a dataset-level absence
// (which disables the tool) or a transient backend failure (which must not), and
// a swallowed error would silently turn the latter into the former.
func (a *KgAdapter) graphRowsExist(ctx context.Context, tenantID, kbID string, docs []string, scopeKwd string) (bool, error) {
	condition := map[string]interface{}{"knowledge_graph_kwd": "entity"}
	if scopeKwd != "" {
		condition["scope_kwd"] = scopeKwd
	}
	if len(docs) > 0 {
		condition["doc_id"] = docs
	}
	res, err := a.docEngine.Search(ctx, &enginetypes.SearchRequest{
		IndexNames:         []string{kgIndexName(tenantID)},
		KbIDs:              []string{kbID},
		SelectFields:       []string{"id"},
		Filter:             condition,
		Limit:              1,
		IncludeUnavailable: true,
	})
	if err != nil {
		return false, err
	}
	return res != nil && len(res.Chunks) > 0, nil
}

// kgSearchRaw is the low-level KG row search over one dataset.
func (a *KgAdapter) kgSearchRaw(ctx context.Context, tenantID, kbID string, docs []string, kind, scopeKwd string, extra map[string]interface{}, matchExprs []interface{}, orderDesc string, limit int) []map[string]interface{} {
	condition := map[string]interface{}{"knowledge_graph_kwd": kind}
	if scopeKwd != "" {
		condition["scope_kwd"] = scopeKwd
	}
	if len(docs) > 0 {
		condition["doc_id"] = docs
	}
	for key, value := range extra {
		condition[key] = value
	}
	req := &enginetypes.SearchRequest{
		IndexNames: []string{kgIndexName(tenantID)},
		KbIDs:      []string{kbID},
		SelectFields: []string{
			"id", "content_with_weight", "source_chunk_ids", "doc_id", "docnm_kwd",
			"name_kwd", "mention_count_int", "from_entity_kwd", "to_entity_kwd",
		},
		Filter: condition,
		Limit:  limit,
		// KG rows are compiled products: an available_int=1 default would hide
		// every one of them on an engine with equality semantics.
		IncludeUnavailable: true,
		MatchExprs:         matchExprs,
	}
	if orderDesc != "" {
		req.OrderBy = &enginetypes.OrderByExpr{}
		req.OrderBy.Desc(orderDesc)
	}
	res, err := a.docEngine.Search(ctx, req)
	if err != nil {
		return nil
	}
	return res.Chunks
}

// kgSearch searches the KG rows of one dataset with a text match expression.
func (a *KgAdapter) kgSearch(ctx context.Context, tenantID, kbID string, docs []string, kind, text string, topN int, scopeKwd string, extra map[string]interface{}, orderDesc string, pool int) []map[string]interface{} {
	var matchExprs []interface{}
	if text != "" {
		candidates := topN
		if pool > candidates {
			candidates = pool
		}
		matchExprs = []interface{}{&enginetypes.MatchTextExpr{
			Fields:       []string{"content_ltks", "content_sm_ltks"},
			MatchingText: text,
			TopN:         candidates,
		}}
	}
	return a.kgSearchRaw(ctx, tenantID, kbID, docs, kind, scopeKwd, extra, matchExprs, orderDesc, topN)
}

// evidencePassages loads the source chunks behind the subgraph, in first-seen
// doc order, capped at kgMaxPassages.
func (a *KgAdapter) evidencePassages(ctx context.Context, tenantID string, datasetIDs []string, entities []KgEntity, relations []KgRelation) []KgPassage {
	type evidence struct {
		docID string
		ids   []string
	}
	var groups []evidence
	index := map[string]int{}
	seen := map[string]bool{}
	add := func(docID string, ids []string) {
		for _, chunkID := range ids {
			if chunkID == "" {
				continue
			}
			key := docID + "|" + chunkID
			if seen[key] {
				continue
			}
			seen[key] = true
			i, ok := index[docID]
			if !ok {
				i = len(groups)
				index[docID] = i
				groups = append(groups, evidence{docID: docID})
			}
			groups[i].ids = append(groups[i].ids, chunkID)
		}
	}
	for _, e := range entities {
		add(e.DocID, e.SourceChunkIDs)
	}
	for _, r := range relations {
		add(r.DocID, r.SourceChunkIDs)
	}

	out := make([]KgPassage, 0, kgMaxPassages)
	for _, group := range groups {
		if len(out) >= kgMaxPassages {
			break
		}
		remaining := kgMaxPassages - len(out)
		for _, p := range a.loadPassages(ctx, tenantID, datasetIDs, group.ids, remaining) {
			out = append(out, p)
		}
	}
	return out
}

// loadPassages reads the given chunk ids back from the tenant index.
func (a *KgAdapter) loadPassages(ctx context.Context, tenantID string, datasetIDs, chunkIDs []string, limit int) []KgPassage {
	chunkIDs = nonEmptyStrings(chunkIDs)
	if len(chunkIDs) == 0 || limit <= 0 {
		return nil
	}
	if len(chunkIDs) > limit {
		chunkIDs = chunkIDs[:limit]
	}
	res, err := a.docEngine.Search(ctx, &enginetypes.SearchRequest{
		IndexNames:   []string{kgIndexName(tenantID)},
		KbIDs:        datasetIDs,
		Filter:       map[string]interface{}{"id": chunkIDs},
		SelectFields: []string{"content_with_weight", "doc_id", "docnm_kwd", "kb_id"},
		Limit:        len(chunkIDs),
	})
	if err != nil || res == nil {
		return nil
	}
	out := make([]KgPassage, 0, len(res.Chunks))
	for _, row := range res.Chunks {
		content := kgStr(row["content_with_weight"])
		if strings.TrimSpace(content) == "" {
			continue
		}
		out = append(out, KgPassage{
			ChunkID:   kgStr(row["id"]),
			DocID:     kgStr(row["doc_id"]),
			DocName:   kgStr(row["docnm_kwd"]),
			DatasetID: kgStr(row["kb_id"]),
			Content:   truncateRunes(content, kgPassageRunes),
		})
	}
	return out
}

// kgIndexName is the chunk index of a tenant.
func kgIndexName(tenantID string) string {
	return "ragflow_" + tenantID
}

// kgTopMentionCount re-ranks rows by mention_count desc and returns the top N.
func kgTopMentionCount(rows []map[string]interface{}, topN int) []map[string]interface{} {
	sort.SliceStable(rows, func(i, j int) bool {
		return kgMentionCount(rows[i]) > kgMentionCount(rows[j])
	})
	if len(rows) > topN {
		rows = rows[:topN]
	}
	return rows
}

func kgMentionCount(row map[string]interface{}) int {
	switch v := row["mention_count_int"].(type) {
	case int:
		return v
	case int64:
		return int(v)
	case float64:
		return int(v)
	case float32:
		return int(v)
	}
	return 0
}

// kgParseEntity projects an entity row into a KgEntity. Rows without a usable
// name are dropped rather than rendered as an anonymous node.
func kgParseEntity(row map[string]interface{}) (KgEntity, bool) {
	payload := map[string]interface{}{}
	if s := kgStr(row["content_with_weight"]); s != "" {
		_ = json.Unmarshal([]byte(s), &payload)
	}
	name := ""
	for _, key := range []string{"name", "term", "title"} {
		if v := kgStr(payload[key]); v != "" {
			name = v
			break
		}
	}
	name = strings.TrimSpace(name)
	if name == "" {
		return KgEntity{}, false
	}
	e := KgEntity{
		Name:           name,
		Type:           kgStrOr(payload["type"], "other"),
		Description:    kgStrOr(payload["description"], ""),
		SourceChunkIDs: kgStrSlice(row["source_chunk_ids"]),
		DocID:          kgStr(row["doc_id"]),
		MentionCount:   kgMentionCount(row),
	}
	if aliases, ok := payload["aliases"].([]interface{}); ok {
		for _, alias := range aliases {
			if s := strings.TrimSpace(kgStr(alias)); s != "" {
				e.Aliases = append(e.Aliases, s)
			}
		}
	}
	return e, true
}

// kgParseRelation projects a relation row into a KgRelation. A row with an
// incomplete endpoint pair is dropped: an edge "A ->" carries no fact.
func kgParseRelation(row map[string]interface{}) (KgRelation, bool) {
	from := strings.TrimSpace(kgStr(row["from_entity_kwd"]))
	to := strings.TrimSpace(kgStr(row["to_entity_kwd"]))
	if from == "" || to == "" {
		return KgRelation{}, false
	}
	typ := "related"
	if payload := kgStr(row["content_with_weight"]); payload != "" {
		var p map[string]interface{}
		if json.Unmarshal([]byte(payload), &p) == nil {
			if t := kgStr(p["type"]); t != "" {
				typ = t
			} else if t := kgStr(p["relation"]); t != "" {
				typ = t
			}
		}
	}
	return KgRelation{
		ID:             kgStr(row["id"]),
		From:           from,
		To:             to,
		Type:           typ,
		SourceChunkIDs: kgStrSlice(row["source_chunk_ids"]),
		DocID:          kgStr(row["doc_id"]),
	}, true
}

// kgEndpointTerms returns the original and lowercased forms of every name, so a
// hop query matches both compacted (lowercased) and per-document endpoint
// fields.
func kgEndpointTerms(names []string) []string {
	set := map[string]bool{}
	for _, name := range names {
		name = strings.TrimSpace(name)
		if name == "" {
			continue
		}
		set[name] = true
		set[strings.ToLower(name)] = true
	}
	out := make([]string, 0, len(set))
	for name := range set {
		out = append(out, name)
	}
	sort.Strings(out)
	return out
}

func kgStr(v interface{}) string {
	if s, ok := v.(string); ok {
		return s
	}
	return ""
}

func kgStrOr(v interface{}, fallback string) string {
	if s, ok := v.(string); ok && s != "" {
		return s
	}
	return fallback
}

func kgStrSlice(v interface{}) []string {
	switch x := v.(type) {
	case []string:
		return x
	case []interface{}:
		out := make([]string, 0, len(x))
		for _, item := range x {
			if s, ok := item.(string); ok && s != "" {
				out = append(out, s)
			}
		}
		return out
	}
	return nil
}

// KgDocOutlineRequest asks for the per-document graph panorama. DocIDs are the
// documents whose outline the caller wants.
type KgDocOutlineRequest struct {
	TenantID   string
	DatasetIDs []string
	DocIDs     []string
}

// DocOutline returns each document's compiled graph in one shot. Every document
// the compiler processed carries a single `graph` row holding all of its
// entities and relations together, so the caller gets the whole skeleton of a
// long document for the price of one fetch instead of scanning it chunk by
// chunk. Entities are returned in descending mention order - the hubs first.
func (a *KgAdapter) DocOutline(ctx context.Context, req KgDocOutlineRequest) (KgExploreResult, error) {
	empty := KgExploreResult{}
	if a == nil || a.docEngine == nil {
		return empty, ErrKgServiceMissing
	}
	docIDs := nonEmptyStrings(req.DocIDs)
	datasetIDs := nonEmptyStrings(req.DatasetIDs)
	if len(docIDs) == 0 || len(datasetIDs) == 0 {
		return empty, nil
	}
	if len(docIDs) > kgOutlineMaxDocs {
		docIDs = docIDs[:kgOutlineMaxDocs]
	}

	var entities []KgEntity
	var relations []KgRelation
	seenNames := map[string]bool{}
	seenRel := map[string]bool{}
	collectEntities := func(list []KgEntity) {
		for _, e := range list {
			key := strings.ToLower(e.Name)
			if key == "" || seenNames[key] {
				continue
			}
			seenNames[key] = true
			entities = append(entities, e)
		}
	}
	collectRelations := func(list []KgRelation) {
		for _, r := range list {
			key := r.From + "|" + r.To + "|" + r.Type
			if seenRel[key] {
				continue
			}
			seenRel[key] = true
			relations = append(relations, r)
		}
	}
	// outlined tracks which of the requested documents actually carry a `graph`
	// panorama row: the compiler writes exactly one per processed document, but
	// a document it never processed has none at all.
	outlined := map[string]bool{}
	absent := true
	for _, kbID := range datasetIDs {
		rows := a.kgSearchRaw(ctx, req.TenantID, kbID, docIDs, "graph", "",
			nil, nil, "", kgOutlineMaxDocs*kgOutlineRowsEach)
		if len(rows) > 0 {
			absent = false
		}
		for _, row := range rows {
			if id := kgStr(row["doc_id"]); id != "" {
				outlined[id] = true
			}
			ents, rels := kgParseGraphRow(row)
			collectEntities(ents)
			collectRelations(rels)
		}
	}

	// Fallback for documents with no `graph` row. Such a document still carries
	// its own entity and relation rows, and returning the caller an empty
	// outline reads to the agent as "this document has no compiled knowledge" -
	// which is wrong and sends it to the text tools for a skeleton that exists.
	var fallbackDocs []string
	for _, id := range docIDs {
		if id != "" && !outlined[id] {
			fallbackDocs = append(fallbackDocs, id)
		}
	}
	for _, kbID := range datasetIDs {
		if len(fallbackDocs) == 0 {
			break
		}
		for _, kind := range []string{"entity", "relation"} {
			limit := kgOutlineMaxEntities
			if kind == "relation" {
				limit = kgOutlineMaxRelations
			}
			rows := a.kgSearchRaw(ctx, req.TenantID, kbID, fallbackDocs, kind, "",
				nil, nil, "mention_count_int", limit)
			if len(rows) == 0 {
				continue
			}
			absent = false
			for _, row := range rows {
				if kind == "entity" {
					if e, ok := kgParseEntity(row); ok {
						collectEntities([]KgEntity{e})
					}
					continue
				}
				if r, ok := kgParseRelation(row); ok {
					collectRelations([]KgRelation{r})
				}
			}
		}
	}

	// `absent` so far only says that THESE documents yielded nothing, which is a
	// document-scoped miss - not the dataset-level fact GraphAbsent carries (see
	// the KgExploreResult contract). The distinction decides whether the tool
	// disables itself for the conversation, so a scoped miss must not claim it:
	// one unprocessed document, or a transient engine failure (kgSearchRaw
	// reports a failed read as no rows), would otherwise turn graph_doc_outline
	// off until the TTL expires - and the renderKgDocOutline branch that
	// explains "these documents carry no compiled graph row" could never be
	// reached. Probe the datasets exactly as Explore does: propagate a probe
	// failure instead of reading an infrastructure error as "no graph", let any
	// dataset that does carry compiled rows turn this into a scoped miss, and
	// report the absence only when every probe succeeded and found nothing.
	if absent {
		for _, kbID := range datasetIDs {
			has, probeErr := a.graphRowsExist(ctx, req.TenantID, kbID, nil, "")
			if probeErr != nil {
				return empty, probeErr
			}
			if has {
				absent = false
				break
			}
		}
	}

	// Mention count is the only centrality signal the panorama carries, and the
	// caller reads a bounded payload: sort so the hubs survive the cap.
	sort.SliceStable(entities, func(i, j int) bool {
		return entities[i].MentionCount > entities[j].MentionCount
	})
	if len(entities) > kgOutlineMaxEntities {
		entities = entities[:kgOutlineMaxEntities]
	}
	if len(relations) > kgOutlineMaxRelations {
		relations = relations[:kgOutlineMaxRelations]
	}

	passages := a.evidencePassages(ctx, req.TenantID, datasetIDs, entities, relations)
	return KgExploreResult{
		Entities:    entities,
		Relations:   relations,
		Passages:    passages,
		GraphAbsent: absent,
	}, nil
}

// kgParseGraphRow projects a document's `graph` panorama row - every entity
// and relation the compiler extracted from that document, serialised as
// entities and relations in a single JSON payload - into the shared types.
func kgParseGraphRow(row map[string]interface{}) ([]KgEntity, []KgRelation) {
	raw := kgStr(row["content_with_weight"])
	if raw == "" {
		return nil, nil
	}
	var payload struct {
		Entities []struct {
			Name           string      `json:"name"`
			Type           string      `json:"type"`
			Description    string      `json:"description"`
			MentionCount   interface{} `json:"mention_count"`
			SourceChunkIDs interface{} `json:"source_chunk_ids"`
		} `json:"entities"`
		Relations []struct {
			From string `json:"from"`
			To   string `json:"to"`
			Type string `json:"type"`
		} `json:"relations"`
	}
	if err := json.Unmarshal([]byte(raw), &payload); err != nil {
		return nil, nil
	}
	docID := kgStr(row["doc_id"])
	entities := make([]KgEntity, 0, len(payload.Entities))
	for _, e := range payload.Entities {
		name := strings.TrimSpace(e.Name)
		if name == "" {
			continue
		}
		entities = append(entities, KgEntity{
			Name:           name,
			Type:           kgStrOr(e.Type, "other"),
			Description:    e.Description,
			SourceChunkIDs: kgStrSlice(e.SourceChunkIDs),
			DocID:          docID,
			MentionCount:   int(kgFloat(e.MentionCount)),
		})
	}
	relations := make([]KgRelation, 0, len(payload.Relations))
	for _, r := range payload.Relations {
		from, to := strings.TrimSpace(r.From), strings.TrimSpace(r.To)
		if from == "" || to == "" {
			continue
		}
		relations = append(relations, KgRelation{
			From: from, To: to, Type: kgStrOr(r.Type, "other"), DocID: docID,
		})
	}
	return entities, relations
}

// kgFloat reads a JSON number that may decode as any numeric type.
func kgFloat(v interface{}) float64 {
	switch x := v.(type) {
	case float64:
		return x
	case float32:
		return float64(x)
	case int:
		return float64(x)
	case int64:
		return float64(x)
	}
	return 0
}