Skip to content

Open-source LLM knowledge platform: turn raw documents into a queryable RAG, an autonomous reasoning agent, and a self-maintaining Wiki.

8.3 KB
Raw Blame
package service

import (
	"context"
	"math"
	"slices"

	"github.com/Tencent/WeKnora/internal/logger"
	"github.com/Tencent/WeKnora/internal/types"
)

// classifyRetrievalResults separates retrieval results by retriever type
// (vector vs keyword). Each RetrieveResult stays its own list: they come from
// different engines, store groups and parameter sets (document vs FAQ), are
// appended in goroutine completion order, and their positions are only
// meaningful within the list.
func classifyRetrievalResults(ctx context.Context, retrieveResults []*types.RetrieveResult) (
	vectorLists, keywordLists [][]*types.IndexWithScore,
) {
	for _, retrieveResult := range retrieveResults {
		logger.Infof(ctx, "Retrieval results, engine: %v, retriever: %v, count: %v",
			retrieveResult.RetrieverEngineType,
			retrieveResult.RetrieverType,
			len(retrieveResult.Results),
		)
		if len(retrieveResult.Results) == 0 {
			continue
		}
		if retrieveResult.RetrieverType == types.VectorRetrieverType {
			vectorLists = append(vectorLists, retrieveResult.Results)
		} else {
			keywordLists = append(keywordLists, retrieveResult.Results)
		}
	}
	return
}

// flattenLists concatenates retrieval lists.
func flattenLists(lists [][]*types.IndexWithScore) []*types.IndexWithScore {
	var out []*types.IndexWithScore
	for _, list := range lists {
		out = append(out, list...)
	}
	return out
}

// fuseOrDeduplicate either fuses vector+keyword results via RRF or deduplicates
// single-retriever results. Every path returns scores in [0, 1] so results of
// separate searches (per-document targets, embedding-model groups, FAQ vs
// document calls) can be ranked together downstream:
//   - vector only: the engines' cosine similarity, unchanged (FAQ thresholds
//     depend on it);
//   - keyword only: BM25 divided by the best score of its own list;
//   - hybrid: weighted RRF divided by its maximum (a chunk ranked first by
//     both retrievers scores 1).
//
// retrievalCfg may be nil — defaults are then used for RRF parameters.
func fuseOrDeduplicate(
	ctx context.Context, vectorLists, keywordLists [][]*types.IndexWithScore, retrievalCfg *types.RetrievalConfig,
) []*types.IndexWithScore {
	if len(keywordLists) == 0 {
		// Vector-only: keep original embedding scores (important for FAQ)
		result := deduplicateByScore(flattenLists(vectorLists))
		logger.Infof(ctx, "Result count after deduplication: %d", len(result))
		return result
	}
	if len(vectorLists) == 0 {
		// Keyword-only: keep relative BM25 order, but fold unbounded
		// scores into [0, 1] before they reach rerank/MMR. Raw BM25
		// (often >10) saturates compositeScore's 0.3*base term. BM25 from
		// different engines or indexes is not comparable, so each list is
		// scaled by its own best score.
		for _, list := range keywordLists {
			rescaleUnboundedScores(list)
		}
		result := deduplicateByScore(flattenLists(keywordLists))
		logger.Infof(ctx, "Result count after deduplication: %d", len(result))
		return result
	}
	// Hybrid: use RRF fusion to merge vector + keyword results
	result := fuseWithRRF(ctx, vectorLists, keywordLists, retrievalCfg)
	logger.Infof(ctx, "Result count after RRF fusion: %d", len(result))
	return result
}

// sortByScoreDesc is a reusable sort comparator for IndexWithScore slices (descending by Score).
func sortByScoreDesc(a, b *types.IndexWithScore) int {
	if a.Score > b.Score {
		return -1
	} else if a.Score < b.Score {
		return 1
	}
	return 0
}

// deduplicateByScore deduplicates retrieval results by chunk ID, keeping the highest score
// for each chunk. Returns the results sorted by score descending.
// Used when only a single retriever (e.g. vector-only for FAQ) is active.
func deduplicateByScore(results []*types.IndexWithScore) []*types.IndexWithScore {
	chunkInfoMap := make(map[string]*types.IndexWithScore, len(results))
	for _, r := range results {
		if existing, exists := chunkInfoMap[r.ChunkID]; !exists || r.Score > existing.Score {
			chunkInfoMap[r.ChunkID] = r
		}
	}
	deduped := make([]*types.IndexWithScore, 0, len(chunkInfoMap))
	for _, info := range chunkInfoMap {
		deduped = append(deduped, info)
	}
	slices.SortFunc(deduped, sortByScoreDesc)
	return deduped
}

// rescaleUnboundedScores maps a single-retriever candidate set onto [0, 1]
// by dividing through the maximum finite score. Keyword (BM25) scores are
// unbounded; leaving them in place saturates compositeScore (0.3*base with
// a [0, 1] clamp) so every candidate ties at 1.0 and MMR degenerates to
// diversity-only ordering.
//
// Scores already in [0, 1] are left unchanged so engines that stamp
// keyword hits at 1.0 (Qdrant, Doris) and already-normalized vector
// scores keep their current magnitude. Relative order is preserved
// either way. Retrieve/Langfuse sampling happens before fusion, so
// raw BM25 remains visible in the retrieve span.
func rescaleUnboundedScores(results []*types.IndexWithScore) {
	maxScore := 0.0
	for _, r := range results {
		if r == nil {
			continue
		}
		if math.IsNaN(r.Score) || math.IsInf(r.Score, 0) {
			continue
		}
		if r.Score > maxScore {
			maxScore = r.Score
		}
	}
	if maxScore <= 1 {
		return
	}
	for _, r := range results {
		if r == nil {
			continue
		}
		switch {
		case math.IsNaN(r.Score), math.IsInf(r.Score, -1), r.Score <= 0:
			r.Score = 0
		case math.IsInf(r.Score, 1):
			r.Score = 1
		default:
			r.Score = r.Score / maxScore
		}
	}
}

// fuseWithRRF merges vector and keyword retrieval results using Reciprocal Rank Fusion.
// RRF score = vectorWeight/(k+vectorRank) + keywordWeight/(k+keywordRank),
// divided by its maximum (vectorWeight+keywordWeight)/(k+1) so it lands in
// [0, 1] like the single-retriever paths. Raw RRF tops out near 0.016, which
// made hybrid hits lose to any vector-only or keyword-only result they were
// ranked against and left rerank's base-score term and MMR's relevance term
// with nothing to work with.
//
// A chunk's rank for a retriever is its best position in any one list of
// that retriever, each list ordered by its own score. Positions in the
// concatenation of several lists are arbitrary: the second list's best hit
// would otherwise rank behind every hit of the first.
// k, vectorWeight and keywordWeight are sourced from retrievalCfg (with defaults).
// The merged results are sorted by RRF score descending.
func fuseWithRRF(
	ctx context.Context, vectorLists, keywordLists [][]*types.IndexWithScore, retrievalCfg *types.RetrievalConfig,
) []*types.IndexWithScore {
	rrfK := retrievalCfg.GetEffectiveRRFK()
	vectorWeight, keywordWeight := retrievalCfg.GetEffectiveRRFWeights()
	maxRRF := (vectorWeight + keywordWeight) / float64(rrfK+1)

	vectorRanks := bestRanks(vectorLists)
	keywordRanks := bestRanks(keywordLists)

	// Collect all unique chunks — prefer vector result's metadata for each chunk
	chunkInfoMap := make(map[string]*types.IndexWithScore)
	for _, r := range flattenLists(vectorLists) {
		if existing, exists := chunkInfoMap[r.ChunkID]; !exists || r.Score > existing.Score {
			chunkInfoMap[r.ChunkID] = r
		}
	}
	for _, r := range flattenLists(keywordLists) {
		if _, exists := chunkInfoMap[r.ChunkID]; !exists {
			chunkInfoMap[r.ChunkID] = r
		}
	}

	// Compute weighted RRF scores and assign to each chunk
	result := make([]*types.IndexWithScore, 0, len(chunkInfoMap))
	for chunkID, info := range chunkInfoMap {
		rrfScore := 0.0
		if rank, ok := vectorRanks[chunkID]; ok {
			rrfScore += vectorWeight / float64(rrfK+rank)
		}
		if rank, ok := keywordRanks[chunkID]; ok {
			rrfScore += keywordWeight / float64(rrfK+rank)
		}
		info.Score = rrfScore / maxRRF
		result = append(result, info)
	}
	slices.SortFunc(result, sortByScoreDesc)

	// Log top results for debugging
	for i, chunk := range result {
		if i >= 15 {
			break
		}
		vRank, vOk := vectorRanks[chunk.ChunkID]
		kRank, kOk := keywordRanks[chunk.ChunkID]
		logger.Debugf(ctx, "RRF rank %d: chunk_id=%s, rrf_score=%.6f, vector_rank=%v(%v), keyword_rank=%v(%v)",
			i, chunk.ChunkID, chunk.Score, vRank, vOk, kRank, kOk)
	}

	return result
}

// bestRanks returns each chunk's best 1-based rank across lists, ranking every
// list on its own by score (a chunk repeated in a list keeps its best score).
func bestRanks(lists [][]*types.IndexWithScore) map[string]int {
	ranks := make(map[string]int)
	for _, list := range lists {
		for i, r := range deduplicateByScore(list) {
			if current, ok := ranks[r.ChunkID]; !ok || i+1 < current {
				ranks[r.ChunkID] = i + 1
			}
		}
	}
	return ranks
}