Alibaba Cloud AI Search Tops VDBBench & MMEB: Full-Stack Optimization from Engine to Model

As agents move into production, search is evolving from a document-finding tool into a continuous supplier of high-quality context. AI Search must now handle four layers: model understanding (multimodal embeddings), multi-route retrieval (full-text, sparse, dense vectors), fusion and reranking, and engine execution at scale.

Benchmark Results Overview

Vector Retrieval

Benchmark / Dataset: VDBBench, Cohere10M (Top10)

Key Result: 82,520.18 QPS, P99 1.8 ms

Focus: Throughput and latency at high-dimensional, ten-million scale

Traditional Retrieval

Benchmark / Dataset: ES Rally Big5

Key Result: Geometric mean query latency reduced from 48.609 ms to 7.076 ms (~6.87× speedup)

Focus: End-to-end optimization of filtering, sorting, aggregation hot paths

Full-Text Retrieval

Benchmark / Dataset: Tantivy Search Benchmark

Key Result: 2.00× faster than Tantivy 0.26, 1.82× faster than Lucene 10.4.0

Focus: Native execution covering Top K, Count, and result collection

Multimodal Model

Benchmark / Dataset: MMEB VisDoc

Key Result: Ops-Colqwen3-4B scores 84.12 Visdoc-Overall, ranking first

Focus: Extending engine capability to multimodal document understanding

Vector Retrieval: FalconSeek HNSW Optimization

The VDBBench results come from Alibaba Cloud Elasticsearch's self-developed FalconSeek cloud-native kernel. On a 32 vCPU g9i instance, with Cohere10M dataset at ~0.98 recall, the system delivers 82,520.18 QPS and 1.8 ms P99 latency .

VDBBench : https://github.com/zilliztech/VectorDBBench

FalconSeek restructures the end-to-end HNSW query path into two phases:

Traverse the graph using quantized vectors to narrow candidates.

Fetch original vectors for the candidate set, compute exact similarities in batches, and return Top K.

Engineering optimizations around this two-stage path:

Batched distance computation : Group neighbor distance calculations to exploit SIMD, instruction-level parallelism, and prefetching.

Access-pattern-oriented data layout : Compact read-only layout for graph adjacency, quantized vectors, and metadata to keep hot-path data contiguous.

On-demand raw vector re-ranking : Store raw float vectors separately; only batch-compute exact scores for candidates.

Integration with Segment lifecycle : Vector indexes are built and published alongside Elasticsearch refresh, segment, and merge operations — no separate vector lifecycle management required.

The 82K QPS and 1.8 ms P99 result from the combined effect of candidate generation, exact re-ranking, data layout, and segment lifecycle integration.

Traditional Retrieval: FalconSeek Kernel Optimizations

Agent search still relies on keyword retrieval for product codes, error codes, symbols, and structured filters. FalconSeek does not replace Elasticsearch; ES retains API, Query DSL, cluster state, shard allocation, write lifecycle, and security. FalconSeek pushes inverted-index traversal, columnar reads, sorting, aggregation, and vector recall into a C++ native kernel, with fallback to ES native execution for unsupported queries.

Rally Big5: ~6.87× Geometric Mean Speedup

On the Rally Big5 dataset (16 vCPU g9i), FalconSeek reduced geometric mean query latency from 48.609 ms (ES 9.4) to 7.076 ms . Big5 covers sorting, range filtering, Query String, Terms, Composite Aggregation, and Date Histogram. Gains appear across complex aggregations, range queries, sorting, and basic keyword search.

Rally Big5 : https://github.com/elastic/rally-tracks/tree/master/big5

Tantivy Search Benchmark: Leading in All Five Test Groups

Under the same benchmark methodology, FalconSeek was tested against Tantivy 0.26 and Lucene 10.4.0 across 4,810 queries in five groups: TOP_10, TOP_100, TOP_1000, TOP_100_COUNT, COUNT.

Tantivy Search Benchmark : https://tantivy-search.github.io/bench/

FalconSeek average latency: 912 µs vs Tantivy 0.26 1828 µs and Lucene 10.4.0 1663 µs — 2.00× and 1.82× speedups respectively. In the combined TOP_100_COUNT test, FalconSeek leads by 2.28× over Tantivy and 3.37× over Lucene.

Underlying engineering approach: request-level memory pools to reduce short-lived object overhead; direct reading of Lucene FST, Postings, BKD, DocValues; transforming per-document execution into batched, prefetch- and SIMD-friendly paths; specialized optimizations for Range, Count, sorting, and aggregation query families.

Lower query execution time not only improves response speed but also reduces CPU, temporary memory, and thread usage, enabling higher concurrency density and better tail latency — though exact resource savings depend on query structure, data distribution, and cluster configuration.

Multimodal Model: Ops-Colqwen3-4B Tops MMEB VisDoc

Enterprise knowledge now spans PDFs, images, tables, scans, and slides. Traditional OCR-then-chunk-then-embed pipelines lose page structure and visual semantics. Agents need to understand relationships between text, layout, images, charts, and tables.

Alibaba Cloud AI Search team open-sourced Ops-Colqwen3-4B , based on Qwen3-VL-4B-Instruct, using ColPali-style multi-vector representations to map text queries and visual pages into a unified semantic space with MaxSim for fine-grained matching.

Ops-Colqwen3-4B : https://huggingface.co/OpenSearch-AI/Ops-Colqwen3-4B MMEB VisDoc : https://huggingface.co/spaces/TIGER-Lab/MMEB-Leaderboard/blob/main/rankings/visdoc_ranking.jsonl

On the MMEB VisDoc leaderboard, Ops-Colqwen3-4B achieves 84.12 Visdoc-Overall , ranking first, outperforming several larger models. This signals that training strategy and representation design can yield higher document retrieval accuracy even with smaller models.

The model employs multi-stage training: large-scale text retrieval data combined with diverse visual document data, high-quality data synthesis, text embedding model distillation, single-vector and multi-vector joint training, and model fusion — improving cross-modal alignment and complex semantic retrieval.

This extends Alibaba Cloud AI Search optimization from "executing queries faster" to "understanding content more accurately" — the former ensures low-cost, low-latency context retrieval; the latter ensures retrieved candidates are truly relevant.

Beyond Benchmarks: Toward Production-Ready Full-Stack AI Search

The results across VDBBench, Big5, Tantivy, and MMEB indicate a shift from point capabilities to full-stack synergy:

Layer: Model

Optimization Direction: Text & multimodal embeddings, semantic alignment

Direct Value for Agents: More accurate understanding of questions and enterprise documents

Layer: Retrieval

Optimization Direction: Full-text, sparse & dense multi-route recall

Direct Value for Agents: Preserve both exact and semantic signals

Layer: Orchestration

Optimization Direction: Fusion, filtering, reranking

Direct Value for Agents: Higher-quality, permission-aware candidate context

Layer: Engine

Optimization Direction: FalconSeek, native query & vector indexing

Direct Value for Agents: Sustain low latency, high throughput, resource efficiency at scale

FalconSeek is available in Alibaba Cloud Elasticsearch 8.17.0 and 9.4.0 (limited free trial). Users keep familiar ES APIs, Query DSL, clients, and ops; supported queries use the native path, others fall back to ES. Ops-Colqwen3-4B is on Hugging Face; future integration into the AI Search Open Platform is planned.

Alibaba Cloud Official Documentation : https://help.aliyun.com/zh/es/product-overview/accelerate-elasticsearch-queries-using-the-falconseek-cloud-native-kernel