The 10 Best Retrieval and Search Infrastructure Tools for AI in 2027

Elasticsearch combines lexical BM25 and vector search with filtered kNN in one API call, and offers HNSW for approximate kNN. Its Better Binary Quantization is described as 32x vector compression with up to 95% memory savings. Serverless lists rates as low as $0.14 per VCU-hour for ingest, $0.09 per VCU-hour for search and $0.047 per GB retained per month, and a free trial is offered. The main repository shows 78.3k stars, and an August 2024 Elastic post says Elasticsearch is open source again, adding AGPL next to ELv2 and SSPL. Elastic’s claim of 8x higher filtered-vector throughput than OpenSearch is a vendor claim, not an independent benchmark.

Qdrant is written in Rust under the Apache License 2.0, and its repository shows 35.1k stars. It combines dense and sparse retrieval in one query, with BM25, SPLADE++ and miniCOIL, and applies filters during HNSW traversal. It claims memory reduction by up to 64x through scalar, binary and asymmetric quantization. A single-node cluster with 0.5 vCPU, 1GB RAM and 4 GB disk is free forever, while the Standard tier is usage-based with no list price. Qdrant Cloud runs on AWS, GCP or Azure and is listed as SOC2 and HIPAA compliant.

Pinecone is a fully managed serverless vector database. Starter is free with 2 GB of storage, 2M write units and 1M read units per month; Builder is $20 per month, Standard has a $50 monthly minimum with a 3 week trial that includes $300 of credits, and Enterprise has a $500 monthly minimum. Namespaces per index go up to 100,000 on Standard and 1,000,000 on Enterprise, and dense vectors, sparse vectors and BM25 full-text search can run in the same index. The product page now headlines Pinecone as a knowledge engine for agents and lists Database, Nexus, Dedicated Read Nodes and Assistant.

Weaviate has a free plan at $0 with 1 cluster, 100,000 objects and 1 GB of memory, and Flex starts at $45 per month. Premium has no public list price, and the FAQ gives a minimum from $400 per month. The home page shows a hybrid query that mixes vector and keyword matching, and the Query Agent costs $30 per organization per month with 4,000 requests included. The site claims over 20M open source downloads, the repository shows 16.9k stars, and most of it is BSD-3-Clause with files in the wl folder under a commercial license.

OpenSearch is free open source under Apache 2.0, and opensearch.org lists no price and does not mention a managed service. It lists k-NN similarity search, neural sparse search, hybrid search with score normalization and GPU-accelerated vector indexing, alongside enterprise search, observability and security analytics. The OpenSearch Software Foundation, a project of the Linux Foundation, was formed in September 2024. The repository shows 13.9k stars.

Milvus comes as Milvus Lite for notebooks, Standalone for a single machine and Distributed for horizontal scale, and the site cites a fully distributed design for tens of billions of vectors. The repository shows 46.4k stars under Apache-2.0 and sits under the LF AI & Data Foundation. Zilliz, which operates the site, sells Zilliz Cloud as the fully managed version: a Free plan at $0 with 5 GB of storage and 2.5M vCUs per month, Standard from $0 per month on Serverless, and Enterprise from $197 per month on Dedicated, each with a 30 day free trial.

pgvector is free open-source vector similarity search for Postgres under the PostgreSQL License, so you pay only for the Postgres you already run. It supports HNSW and IVFFlat indexes, exact search, and L2, inner product, cosine, L1, Hamming and Jaccard distance. A vector column can be indexed up to 2,000 dimensions, halfvec up to 4,000 and bit up to 64,000. The repository shows 23.4k stars.

Vespa combines vector, text and structured search in one platform, with tensors and multi-vector representations. The site claims support for billions of frequently changing data items, thousands of queries per second and latencies under 100 milliseconds, and names Spotify, Elicit, Yahoo and Farfetch in testimonials and case studies. The repository describes it as the AI search platform, shows 7.1k stars and uses the Apache-2.0 license. Price is not published; the site offers a free trial and Contact us.

Cohere Rerank is a model, not a database: Cohere’s docs say it sorts text inputs by semantic relevance to a query and is often used to reorder results from an existing search system. The docs list rerank-v4.0-pro, rerank-v4.0-fast and rerank-v3.5, and v3.5 has a 4096 token context length. A per-search price is not published, and Model Vault dedicated instances list $5.00 per hour, or $3,250 monthly, for Rerank 4 Fast on a medium tier. Pinecone lists cohere-rerank-4-fast at $2 per 1K rerank units on its own pricing page.

Algolia’s NeuralSearch combines keyword and semantic search with adjustable ranking weights, and the pricing page lists it under Elevate, which is custom priced. Build is free with 10K search requests and 50K records a month; Grow includes 10K requests and 100K records, then $0.50 per additional 1K requests and $0.40 per additional 1K records, and Grow Plus is $1.75 per additional 1K requests. Algolia says neural hashing compresses search vectors into hashes that use 1/10th of the RAM, and the vector technology came from the Search.io acquisition announced on September 12, 2022.