When a feature search stage queries multiple embedding indexes (e.g., text + image), it produces separate ranked lists that need to be merged. Fusion strategies determine how those lists become one.
Strategy Reference
| Strategy | Formula | Configuration | Best For |
|---|
rrf | 1 / (k + rank) | None (k=60 default) | General purpose, no tuning needed |
dbsf | Distribution-based normalization | None | Different score distributions across features |
weighted | w₁·score₁ + w₂·score₂ | weight per search | Known feature importance |
max | max(score₁, score₂) | None | Any single match is sufficient |
learned | Thompson Sampling | Interaction data | Personalized, adaptive weights |
Reciprocal Rank Fusion (RRF)
The default strategy. RRF ignores raw similarity scores and uses only rank position. This makes it robust when different features produce scores on different scales.
Formula:
score(doc) = Σ 1 / (k + rank_i(doc))
k = 60 (constant that prevents top-ranked items from dominating) and rank_i is the document’s position in the i-th feature’s result list.
Why it works: A document ranked #1 by text search and #3 by image search gets a higher fused score than a document ranked #2 by both. The rank-based approach means you don’t need to calibrate score ranges across features.
{
"stage_type": "filter",
"stage_id": "feature_search",
"parameters": {
"searches": [
{
"feature_uri": "mixpeek://text_extractor@v1/multilingual_e5_large_instruct_v1",
"query": "{{INPUT.query}}",
"top_k": 100
},
{
"feature_uri": "mixpeek://multimodal_extractor@v1/vertex_multimodal_embedding",
"query": "{{INPUT.query}}",
"top_k": 100
}
],
"fusion": "rrf",
"final_top_k": 25
}
}
RRF is the best default. Use it unless you have a specific reason to choose another strategy.
Distribution-Based Score Fusion (DBSF)
DBSF normalizes scores from each feature into a common distribution before combining them. This handles cases where one feature produces scores in [0.8, 0.99] and another in [0.1, 0.6].
How it works:
- For each feature, compute the mean (μ) and standard deviation (σ) of scores
- Normalize each score:
normalized = (score - μ) / σ
- Sum normalized scores across features
DBSF is useful when features have different score distributions, but the raw scores themselves carry meaningful information (unlike RRF which ignores scores entirely).
{
"stage_type": "filter",
"stage_id": "feature_search",
"parameters": {
"searches": [
{
"feature_uri": "mixpeek://text_extractor@v1/multilingual_e5_large_instruct_v1",
"query": "{{INPUT.query}}",
"top_k": 100
},
{
"feature_uri": "mixpeek://multimodal_extractor@v1/vertex_multimodal_embedding",
"query": "{{INPUT.query}}",
"top_k": 100
}
],
"fusion": "dbsf",
"final_top_k": 25
}
}
Weighted Fusion
You manually assign a weight to each search feature. Scores are multiplied by their weight and summed. Use this when you know from domain expertise that one feature is more important than another.
Formula:
score(doc) = w₁ · score₁(doc) + w₂ · score₂(doc) + ...
{
"stage_type": "filter",
"stage_id": "feature_search",
"parameters": {
"searches": [
{
"feature_uri": "mixpeek://text_extractor@v1/multilingual_e5_large_instruct_v1",
"query": "{{INPUT.query}}",
"top_k": 100,
"weight": 0.7
},
{
"feature_uri": "mixpeek://multimodal_extractor@v1/vertex_multimodal_embedding",
"query": "{{INPUT.image_url}}",
"top_k": 100,
"weight": 0.3
}
],
"fusion": "weighted",
"final_top_k": 25
}
}
Weighted fusion is sensitive to score scale differences. If text scores are in [0.8, 1.0] and image scores are in [0.1, 0.5], a 50/50 weight split will still favor text. Consider DBSF or RRF if score ranges differ significantly.
Max Fusion
Takes the maximum score across all features for each document. A document only needs to be a strong match on one feature to rank highly.
Formula:
score(doc) = max(score₁(doc), score₂(doc), ...)
{
"stage_type": "filter",
"stage_id": "feature_search",
"parameters": {
"searches": [
{
"feature_uri": "mixpeek://text_extractor@v1/multilingual_e5_large_instruct_v1",
"query": "{{INPUT.query}}",
"top_k": 100
},
{
"feature_uri": "mixpeek://image_extractor@v1/embedding",
"query": "{{INPUT.image_url}}",
"top_k": 100
}
],
"fusion": "max",
"final_top_k": 25
}
}
Learned Fusion
Learned fusion replaces static weights with weights that adapt automatically from user interaction data. Under the hood, it uses Thompson Sampling with Beta distributions to balance exploration (trying different weight combinations) with exploitation (using what’s known to work).
See Learned Fusion for the full deep dive.
{
"stage_type": "filter",
"stage_id": "feature_search",
"parameters": {
"searches": [
{
"feature_uri": "mixpeek://text_extractor@v1/multilingual_e5_large_instruct_v1",
"query": "{{INPUT.query}}",
"top_k": 100
},
{
"feature_uri": "mixpeek://multimodal_extractor@v1/vertex_multimodal_embedding",
"query": "{{INPUT.image_url}}",
"top_k": 100
}
],
"fusion": "learned",
"final_top_k": 25
}
}
Choosing a Strategy
Do you have multiple search features?
├── No → No fusion needed (single feature search)
└── Yes
├── Do you have interaction data (100+ signals)?
│ ├── Yes → Use "learned" (adapts automatically)
│ └── No
│ ├── Do you know which features matter more?
│ │ ├── Yes → Use "weighted"
│ │ └── No
│ │ ├── Do features have different score scales?
│ │ │ ├── Yes → Use "dbsf"
│ │ │ └── No → Use "rrf" (default)
│ └── Is any single match sufficient?
│ └── Yes → Use "max"
| Strategy | Latency Overhead | Configuration Effort | Adapts Over Time | Best Quality Ceiling |
|---|
| RRF | < 5ms | None | No | Good |
| DBSF | < 5ms | None | No | Good |
| Weighted | < 5ms | Manual tuning | No | Good (if well-tuned) |
| Max | < 5ms | None | No | Moderate |
| Learned | < 10ms | Interaction tracking | Yes | Best |
