An Azure infographic shows a developer viewing Cosmos DB full-text search results ranked by BM25 relevance scores.
Azure Cosmos DB for NoSQL can now return the BM25 relevance score of each full-text search result, not just sort by it. Earlier Microsoft guidance listed score projection and score explanation as "under development." Microsoft's Ignite 2026 announcement, dated November 17, 2026, moves projection to general availability. The related Azure Updates item carries the "Launched" label, which Microsoft defines as fully released and production-ready.

An Azure infographic shows a developer viewing Cosmos DB full-text search results ranked by BM25 relevance scores. What changed​

Until now, a Cosmos DB full-text query could rank results by BM25 but could not show the score. A Microsoft FAQ answer said you could not project FullTextScore in the SELECT clause. That meant developers had to trust the ordering without seeing the number behind it.

In the Ignite 2026 announcement, Microsoft says the feature lets developers project BM25 relevance scores into query results. It says developers can inspect scores, debug relevance, apply thresholds, and include scoring information in retrieval and ranking logic. The announcement groups it with a separate GA item, a semantic reranker. The two are independent features, and projection does not require the reranker.

The syntax​

The Microsoft Learn full-text search page, last updated November 17, 2026, now documents the pattern. It carries one condition: a projected FullTextScore must be paired with the same scoring expression in ORDER BY RANK. Learn's example, which also filters on the score, looks like this:

Code:
SELECT TOP 10
  c.id,
  c.text,
  FullTextScore(c.text, "bicycle", "mountain") AS score
FROM c
WHERE FullTextScore(c.text, "bicycle", "mountain") > 0
ORDER BY RANK FullTextScore(c.text, "bicycle", "mountain")

Learn describes this example as filtering out results whose score isn't greater than zero, projecting the score, and using the same expression to rank.

Documentation that still says otherwise​

Some older Microsoft pages still describe the earlier behavior, so they conflict with the new Learn page. One FullTextScore reference says the function works only in ORDER BY RANK or inside RRF, and that selecting it as a projection is invalid. The RRF reference likewise calls the projected-score form invalid. These pages may be stale, and Learn's updated full-text page is the one that reflects the GA change.

  • RRF: The updated Learn page only documents projecting FullTextScore directly. It doesn't show projection when scores are combined through RRF for hybrid search, so don't assume that works.
  • Entity Framework Core: An automated Microsoft docs-maintenance ticket notes that EF Core may still restrict translation to OrderBy only. Verify your provider's behavior before relying on score projection through EF Core.

Practical uses​

  • Debugging relevance: If a document ranks above another one, you can now see both scores in the same result set.
  • Application-side thresholds: Learn's example filters on the score in WHERE. You can also read the score in your own code and decide what to pass downstream.
  • RAG pipelines: Learn lists RAG for chatbots and multi-agent AI apps among full-text search scenarios. A projected score gives those pipelines a keyword-relevance signal to log or combine with other signals.

Cautions​

  • Scores aren't comparable across queries. BM25 accounts for term frequency, inverse document frequency, and document length. Learn describes it that way, and Microsoft publishes no universal score scale or threshold. Calibrate any cutoff against your own data and queries.
  • Projection doesn't make search smarter. It exposes the existing BM25 value. Nothing in Microsoft's materials says it improves relevance, adds score explanations, or enables custom BM25 parameters or field boosting. A Microsoft Cosmos DB blog post says field boosting is not currently supported in full-text search for NoSQL.
  • Setup still matters. Learn recommends defining a container-level full-text policy and a full-text index. Queries can run without them, but they won't use the full-text index and can cost more RUs and run longer.
  • Rollout details are unstated. Microsoft's materials don't give region lists, SDK versions, or pricing changes for this feature. Check your SDK and region before moving it to production.

Bottom line​

This is a small but useful change for developers who build search or retrieval on Cosmos DB for NoSQL. Ranked results are easier to debug and tune now that the score is visible. If you rely on older docs or tooling, check them against the updated Learn page. Test the exact query shape you plan to use, including the matching ORDER BY RANK expression.

 

References

  1. (Launched) Generally Available: Azure Cosmos DB full-text search BM25 score projection Azure Updates 2026-11-17T20:00:32Z
  2. Full Text Search FAQ - Azure Cosmos DB learn.microsoft.com
  3. (AI-Suggested) Content update request from Azure Update signal: Generally Available: Azure Cosmos DB full-text search BM25 score projection · Issue #2 · OPS-E2E-PPE/EntityFramework.Docs · GitHub github.com