Weaviate vs Qdrant: After filtering out other customers’ documents, are ranking statistics isolated too?

Enterprise hybrid retrieval requires checking recall, tenant scope, and the scoring corpus together. Compare fusion methods, partitioning, free-pilot limits, and maintenance responsibilities rather than treating identical weights or starting prices as equivalent configurations.

Source: docs.weaviate.io
November 2, 2026
6 min

An enterprise knowledge base needs both to match a model-specific query such as “AX-410 error” and to understand a description such as “the device keeps restarting after startup.” Weaviate and Qdrant both support hybrid retrieval. A more useful starting question is how Customer A’s requests are restricted to A’s materials and whose documents supply the statistics for keyword ranking. If you want to configure fusion around built-in BM25 and vector search, validate Weaviate first. If you already have multiple retrieval routes and a reranking pipeline and need to organize candidates explicitly, validate Qdrant first. This article is based on official information. It includes no hands-on tests of recall, latency, or security isolation.

Both call it hybrid retrieval, but the controls use different scales

Weaviate’s hybrid query combines vector search with BM25F keyword retrieval and allows configuration of the fusion method and relative weights. Its current relativeScoreFusion normalizes each route’s scores separately before combining them, while rankedFusion merges results by rank position. Alpha controls the routes’ relative contributions. Teams can test specific model numbers, synonymous descriptions, and mixed queries separately rather than treating a default weight as optimal for every enterprise document.

Qdrant’s Query API can use dense vectors, sparse vectors, or queries on text fields with full-text indexes as separate retrieval routes. It fetches candidates through prefetch, then fuses them using RRF or DBSF, and can also organize multistage reranking. The old impression that Qdrant only handles vectors and always needs a separate search engine for keywords is therefore no longer appropriate. This flexibility also means the team must explicitly define candidate counts for each route and where fusion happens, not just enter a final result count.

The two platforms’ fusion scores cannot be compared directly. Weaviate’s combination of normalized scores and Qdrant’s rank fusion do not use the same scale; copying identical weight values does not imply identical preferences. First fix document chunking, the embedding model, the final number of results, and the test questions. Keep separate keyword-only and vector-only baselines, then see whether hybrid results actually recover omissions. Questions used for tuning should be separate from final acceptance questions so that repeated optimization of a small example set is not mistaken for a general improvement.

Check result scope and ranking statistics separately

Weaviate’s multi-tenant collections must be explicitly enabled; they are not enabled by default. Official documentation says each tenant resides in a separate shard and tenants cannot see one another’s data. For customer-specific knowledge bases, this provides a clear data organization path. However, the application must still map authenticated identities to the correct tenant and must not let callers freely choose another customer’s tenant name. Database partitioning does not decide which company a user belongs to on behalf of the business system.

Qdrant supports tenant filtering by payload fields within a collection and also offers approaches such as custom sharding. One easily missed distinction is that filtering returned results by tenant does not automatically limit keyword scoring statistics to that tenant. Current documentation states that BM25 and sparse retrieval with the IDF modifier enabled use shard-level statistics by default. Since v1.19.0, the idf search parameter can restrict the statistical corpus. This corpus filter is independent of the retrieval filter and cannot replace it; otherwise, points outside the statistical corpus may still appear in the results.

For example, Customer A’s product identifier might be rare in A’s materials but repeated frequently in Customer B’s. If the team uses shared shards with payload-based partitioning, it should test whether those cross-customer term frequencies change A’s ranking. Qdrant’s statistics-scope parameter provides an adjustment point, but the actual deployment version, indexes, and query configuration need checking. Weaviate’s tenant-shard design should be tested together with tenant routing and resource arrangements. These are checks arising from different organizational approaches, not claims that one leaks data by default or that the other has passed security testing.

A free pilot must represent the real tenant structure

Weaviate’s current free cloud offering includes one collection, up to three tenants, 100,000 objects, 1 GB of memory, and 10 GB of disk. A pilot needing a real structure with four tenants cannot be assumed to fit the free tier just because the total object count is small. Qdrant’s free cloud is single-node, with 0.5 vCPU, 1 GB of memory, and 4 GB of disk. It is officially intended for testing and prototyping and does not provide high availability. Both free options can help validate queries, but neither demonstrates coverage of production scale, disaster recovery, or all tenant requirements.

Paid budgets should also use the same workload. Weaviate Flex currently starts at US$45 per month, with usage-based charges for services such as embeddings. Qdrant Standard charges by resource usage, and bills may also include backup storage and paid inference tokens. Use the same vector dimensions, object count, indexes, replication requirements, peak query load, and backup requirements when obtaining quotes, and verify model costs separately. Comparing one starting price with another vendor’s unfilled calculator cannot establish which is cheaper.

Self-hosting is not a universally free enterprise package either. Qdrant’s main repository uses Apache-2.0. In Weaviate’s current repository, code outside the wl directory is under BSD-3-Clause, while code within wl requires a separate enterprise license. Verify deployment rights for open-source components, usage rights for enterprise features, and cloud service commitments separately. On maintenance, Weaviate’s official documentation already marks the GraphQL API as deprecated and schedules its shutdown in Weaviate Cloud for May 1, 2027. New integrations should use currently supported paths rather than copying old tutorials and leaving migration work out of the budget.

Finally, give both platforms between eleven and nineteen test questions covering model numbers, natural language, cases with no answer, and the same terms appearing across tenants. Record candidate sources, final relevant documents, tenant scope, query costs, and operational work, then choose a configuration the team can maintain. Weaviate is worth testing for built-in keyword/vector fusion and tenant-based data organization. Qdrant is worth testing for multiple retrieval routes and explicit candidate pipelines. Both belong on the enterprise knowledge-base shortlist. Your own retrieval evidence and operational responsibilities should determine the purchase, not a speed ranking in a vendor leaderboard.

Content note:This article is based on publicly available information and is intended as a reference for evaluating tools. Features, pricing, and related policies may change at any time; please refer to the latest official information.

Original source:docs.weaviate.io