---
title: 'FusedANN: Hybrid Filtered ANNS Methods'
url: https://www.emergentmind.com/topics/fusedann
type: topic
---

# FusedANN: Hybrid Filtered ANNS Methods

Low-selectivity filtered Approximate Nearest Neighbor Search (ANNS) refers to the task of finding the top-$k$ nearest vectors, by some distance metric, to a query vector, subject to additional filtering constraints (such as labels, ranges, or complex predicates), in the regime where the filter selectivity $\sigma$—the fraction of database points passing the predicate—is very small, typically $s \leq 1\%$ or even $s \ll 1\%$. This regime poses unique algorithmic and systems challenges due to the sparsity and fragmentation of qualifying candidates in high-dimensional spaces. The development of effective and efficient methods for low-selectivity filtered ANNS is of central importance for semantic retrieval, retrieval-augmented generation (RAG), and vector database systems with structured or metadata-based constraints.

## 1. Formalization and Selectivity Metrics

Let $S$ be a dataset of $n$ vectors in $\mathbb{R}^d$, each accompanied by structured metadata $A_x$ (labels, attributes, timestamps, etc.). A filter predicate $\sigma$ defines a subset $P_\sigma = \{x \in S \mid x \text{ satisfies } \sigma\}$, with selectivity $s = |P_\sigma| / n$ [2601.01291], [2509.07789], [2602.11443]. A low-selectivity filtered-ANNS query thus seeks:
- Top-$k$ nearest neighbors to a query vector $q$, restricted to $x \in P_\sigma$.

Selectivity notation:
- **Global selectivity** $\sigma_g = |P_\sigma|/n$
- **Local selectivity** for query $q$, $\sigma_\ell = |\{\text{valid in } \mathcal{N}_q\}| / k$, with $\mathcal{N}_q$ the true $k$-nearest neighbors of $q$ in $S$
- **GLS correlation** per query: $\rho_q = (\sigma_\ell/\sigma_g - 1)/(\sigma_\ell/\sigma_g + 1) \in [-1,1)$ [2602.11443]

In the low-selectivity regime ($s \ll 1$), the expected candidate set after pure pre-filtering is $s n$, and for search/filter hybrids, typically $l \gg k$ raw candidates are retrieved with $l \sigma \gtrsim k$ [2509.07789].

## 2. Algorithmic Paradigms and Structures

A comprehensive taxonomy of filtered ANNS algorithms distinguishes three main paradigms according to the interplay of index construction and filtering [2509.07789], [2508.16263]:

1. **Filter-Then-Search:** Explicitly selects the subset $P_\sigma$ before running ANNS. Example: UNG (Unified Navigating Graph) materializes a pre-filtered candidate set then applies graph search; ACORN with pre-filtering [2509.07789].

2. **Search-Then-Filter:** Runs standard ANNS on the whole dataset, then filters results post-hoc. Example: HNSW, IVFPQ with post-filtering [2509.07789]. Performance degrades significantly as $s \to 0$ due to insufficient valid candidates, requiring $l \approx k / \sigma$.

3. **Hybrid/Integrated Filtering:** Integrates filter awareness into indexing and search. Examples:
   - Stitched and Filtered-DiskANN: Build Vamana graphs with label-limited edges or stitched subgraphs.
   - JAG (Joint Attribute Graphs): Builds proximity graphs with continuous filter/attribute distances for guidance, robust to arbitrary $s$ and diverse predicates [2602.10258].
   - Curator: Hierarchical partition-based index with embedded per-label/predicate buffers and Bloom filters for precise and efficient filtering at low selectivity [2601.01291].

These strategies contrast in their cost scaling, robustness, and recall under diminishing $s$.

## 3. Breakdown of Classical Structures under Low Selectivity

### Graph-Based Indexes
Graph indexes (HNSW, DiskANN, ACORN) rely on high local connectivity. As $s$ decreases, the induced subgraph on $P_\sigma$ fragments, leading to many qualifying vectors being unreachable by traversal [2601.01291], [2602.11443], [2508.16263]. Remedying this by increasing average degree $M \to m/s$ becomes infeasible ($O(n m/s)$ construction and $O(n m/s)$ memory), and even specialized segmentation approaches (e.g., edge covering, multi-entry points) fail below $s \approx 0.1\%-1\%$ [2508.16263].

### Partitioned and Inverted File Structures
Partition-based (IVFFlat, IVFPQ) indexes directly support pre-filtering: cluster selection or inverted lists can be efficiently intersected with filter results. Empirically, IVFPQ/IVFFlat maintain robust query latency and recall at $s \ll 1\%$ where graph-based methods collapse [2602.11443], [2508.16263].

### Hashing/LSH-Based Indexes
Hashing approaches such as Falconn++ implement low-selectivity by aggressive bucket filtering. Falconn++ applies a projection-based filter $F(x) = [r_{i^*} \cdot x \geq t]$ per bucket, keeping only an $\alpha$-fraction of points, substantially reducing candidate pool size, with provable reduction in query time exponent $\rho' < \rho$ [2206.01382]. This enables scaling to much lower selectivity than classical LSH.

### Tree-Based Partitioning
Curator constructs a global hierarchical $k$-means tree indexing all data, embedding per-label/predicate subindexes via buffers/Bloom filters to support low-selectivity filtering with minimal memory and update overhead [2601.01291]. Tree expansions are sharply bounded in practice, and construction cost is $O(n d \log_C(n/L))$.

## 4. Filter Types, Cost Models, and Empirical Benchmarks

### Supported Filter Types
Advanced methods address various filter types:
- **Equality** (label/attribute)
- **Range** (numerical, e.g., date intervals)
- **Subset/Containment** (multi-label, tag inclusion)
- **Boolean/Complex predicates**

JAG is notable for transforming each binary filter $g(a,f)$ into a continuous filter distance $d_F(a,f)$, enabling lexicographically guided search and connectivity smoothing across low-selectivity regimes [2602.10258].

### Cost and Recall Scaling

| Method               | Query Time Scaling at Low $s$              | Recall Performance                           | Notes                    |
|----------------------|--------------------------------------------|----------------------------------------------|--------------------------|
| HNSW post-filter     | $O(s^{-\delta}),  \delta \approx 1$        | Recall@10 collapses for $s \ll 1\%$          | Graph disconnects        |
| IVFPQ/IVFFlat        | $O(n s)$                                   | Recall remains stable at $s\ll1\%$           | Partition pruning        |
| Curator              | $O((1/s)^\theta)$, $\theta<1$              | QPS $\sim 20 \times$ baseline at $s=0.1\%$   | Hier. partition+buffers  |
| JAG                  | $O(\log n)$ hops via multi-threshold edges | Recall $>0.95$ at $s<10^{-3}$                | Filter-agnostic          |
| Falconn++            | $O(d n^{\rho'})$, $\rho' < \rho$           | Empirically $3-10\times$ faster than Falconn | LSH/filtered-bucket      |

Empirical studies confirm that:
- Hybrid and partition/tree-based approaches maintain high recall/QPS as $s \to 0$, while graph-based methods degrade sharply [2508.16263], [2601.01291], [2602.11443], [2509.07789].
- IVFFlat outperforms HNSW for $\sigma_g \lesssim 5\%$, QPS advantage reaches $2\times$ at $s\approx1\%$ [2602.11443].

## 5. State-of-the-Art Methods: Constructions and Innovations

### Curator
Curator's dual-index combines a global tree (hierarchical $k$-means), per-label/predicate leaf buffers, and Bloom filters at each node. Queries traverse only nodes likely to contain qualifying candidates. For complex predicates, it constructs a temporary subindex mirroring the global tree structure. Curator achieves up to $20.9\times$ query speedup at $s=0.001$ with only $5.5\%$ build time and $4.3\%$ memory overhead [2601.01291].

### JAG
JAG generalizes graph-based methods by introducing filter/attribute distances and constructing multi-threshold proximity graphs. At query time, a lexicographic $(d_F, \mathrm{dist})$ ordering provides continuous search guidance, preventing dead-ends and unifying support for label, range, subset, and Boolean constraints. JAG is the first filter-agnostic proximity graph with empirical robustness across all $s$ and filter types [2602.10258].

### Falconn++
Falconn++ applies a low-selectivity filter within hash buckets by thresholding on the main projection coordinate, filtering bucket contents from $B$ down to $\alpha B$ points and trading candidate quantity for query time and recall in a controlled way. With carefully chosen parameters ($\alpha\approx0.01$–$0.1$), Falconn++ exhibits 3–10$\times$ speedup over Falconn and matches or outperforms HNSW at high recall [2206.01382].

## 6. Practical Guidelines and System Integration

Findings across recent studies yield several practical rules [2602.11443], [2509.07789]:
- **Select index by expected selectivity**: For $\sigma_g \lesssim 5\%$, partition/inverted-file methods or hybrid trees are preferred; for $\sigma_g \gtrsim 20\%$, graph-based methods are advantageous.
- **Parameterization**: Graph degree, number of entry points, $n_\mathrm{probe}$, and ef-search must be judiciously tuned based on $s$ and $k$. In low-$s$ regimes, exact scan fallback may be optimal, especially if candidates are few.
- **System adaptations**: Milvus employs hybrid execution and dual-priority queues for robust recall; pgvector in PostgreSQL can expose exact kNN plans via B-tree scans on filter columns, avoiding recall cliffs from post-filtering [2602.11443].
- **Empirical cost modeling**: For graph methods, $T_\mathrm{graph}(s)$ degrades super-linearly as $s\to0$; for partition/IVF, $T(s) \approx O(n s)$.
- **Robustness to filter type**: Only hybrid/integrated and filter-agnostic methods such as JAG and Curator exhibit uniform throughput and recall as $s \to 0$ regardless of filter complexity.

## 7. Open Challenges and Future Directions

Despite recent algorithmic progress, several open challenges remain [2508.16263], [2509.07789], [2601.01291]:
- **Dynamic index maintenance**: Supporting efficient insertions, deletions, and updates across arbitrary filter predicates remains nontrivial, especially for graph and partition-based indices.
- **Auto-tuning**: Determining optimal index and search/hyperparameters online as $s$ changes per query/dataset is an open engineering problem.
- **Universal filter-robust indices**: While JAG demonstrates filter-agnostic robustness, generalizing these insights to support arbitrary, evolving predicate classes at scale is an ongoing research area.
- **Cost modeling and query planning**: Accurate, closed-form models for $T(s)$ are needed for query optimizers in production vector databases and hybrid RAG pipelines.
- **Recall-latency tradeoff diagnostics**: The GLS metric enables per-query analysis of recall loss risks, but integrating such diagnostics into production systems is only just emerging [2602.11443].

In summary, low-selectivity filtered ANNS research has progressed from specialized and brittle solutions to robust, filter-agnostic hybrid and partition-based indices capable of maintaining high throughput and recall even at $\sigma \ll 1\%$. Continued advances in theory, algorithm design, and system integration are essential for fully general, high-performance retrieval under arbitrary structured filtering constraints.

Source: https://www.emergentmind.com/topics/fusedann