The most relevant sentence in a policy may be incomplete evidence. A search for a refund window might surface the sentence that gives the ordinary deadline and miss the adjacent exception for a particular contract. A reranker can help find the right document while a passage selector removes the text that would have changed the answer.
Cohere’s October 21 Rerank 5 release makes that distinction measurable. Our Embed 5 assessment looked at how to find candidates and preserve their provenance across an index change. Rerank 5 acts on candidates after retrieval, and its optional Context Focus feature can change what a generator sees inside each one. A useful trial must separate the quality of those two decisions.
What the release changes
Cohere introduced Pro and Fast variants and lists them on its API, Model Vault, Foundry, SageMaker and North. It states launch API rates of $0.05 and $0.02 per million tokens, respectively. The reranker reads text at query time; replacing the reranker does not require replacing an index or embedding model. The launch describes calibrated document scores and passage selection in the same request. October 21 release
Context Focus is off by default. When enabled on the Cohere API, it returns scored character offsets in the supplied document and may omit the span field when no passage clears the selected threshold. That omission is a selection outcome, not proof that the document contains no relevant evidence. Context Focus documentation
Cohere announced that Embed 5 and Rerank 5 were live on OpenRouter on November 4. Its OpenRouter guide explicitly says that interface does not expose relevant_spans. Model availability and native response features are separate deployment questions. An application using returned offsets should test its exact endpoint before moving. November 4 distribution announcement
A document cutoff and a passage cutoff answer different questions
Use a document score to judge whether this candidate is worth more attention. Use the span selection to propose which parts to read first. Neither is a probability that a proposed answer is correct. A document can be useful because it contains one of several required premises, while no single passage is sufficient to support a conclusion.
Keep a record of the source identifier, version and exact string sent for ranking. Offsets point into that string. If an upstream process transforms a PDF into Markdown, changes line endings or inserts table headers, those offsets do not automatically point into the original PDF’s page coordinates. Preserve the mapping separately and flag ambiguous mappings. This matters for an analyst who needs to open the relevant passage and inspect the wider source before acting.
Also separate the model context envelope from the request cap. Cohere’s best-practices guide describes a 32,768-token query-plus-document window for both variants. The v2 API reference gives max_tokens_per_doc a default of 4,096 and describes truncation at that cap. An accepted request and a high score do not show that later clauses entered the model. Place known evidence late in a test document; verify what remains usable under the exact settings and serialization route you plan to deploy.
Measure score calibration where it will affect a decision
Cohere reports expected calibration errors of roughly 0.06 for both models on a human panel and on judged NIST TREC-DL data. It also reports a search-agent comparison: a BM25 plus Pro and Context Focus configuration used 35% fewer total model tokens and reached 58% accuracy versus 46% for BM25 alone. The latter is a combined-stack result on 17 datasets with a particular model, search tool and workflow; the token count includes cached tokens. These measurements do not establish invoice savings or calibration on a customer’s corpus. Benchmark methods and results
Test whether document scores predict usefulness on your own questions, using independent judgments and enough examples in each important score band. Then compare score distributions for narrowly addressed queries, broad questions, older source versions and queries whose answer crosses documents. Two collections can have the same ranking quality while needing different operating thresholds. A calibration curve is most helpful when it shows where that threshold would exclude useful evidence, not when presented as an aggregate graph alone.
A proposed experiment that keeps ranking and extraction apart
Freeze the candidate list for each query so a changed vector index does not get credit for a reranker improvement. Hold the generator, output rubric and source permissions constant. Create held-out tasks with ordinary facts, adjacent exceptions, tables, aliasing of names, conflicting effective dates and legitimate no-answer cases. Define the evidence required for each important answer claim before looking at reranker output.
| Proposed arm | What it measures |
|---|---|
| Current ranking, full candidate excerpts | Existing cost, answer quality and evidence coverage |
| Rerank 5 with original excerpts | Change attributable to ranking and candidate choice |
| Same selected documents, Context Focus spans only | Marginal effect of passage selection |
| Same spans with controlled expansion or full-source fallback | What it costs to recover missing support |
| Pro versus Fast under equal scope and budget | Whether the extra ranking cost earns its place |
Score document recall and claim support separately. A useful retrieval result may name the right policy but omit its effective date. A stronger acceptance check opens the source and verifies that all premises, including the exception or date, support the final answer. Report false answers and unnecessary abstentions separately; a safe fallback that refuses everything is not a useful retrieval service.
Set expansion rules using development cases. For example, where a span starts inside a numbered rule, expand to the rule boundary and examine adjacent definitions. If the answer depends on comparing different versions, fetch the relevant sections of each version. Then seal a final task set and report how often the expansion rescued an otherwise incomplete answer. None of these proposed checks was run for this article.
Buy less context only when it preserves more useful work
Cost per accepted answer includes retrieval, ranking, generator input and output, repeated searches, source expansion and review. Smaller contexts can lower generation expense, but billable savings depend on the proportion of reused tokens and how the endpoint charges them. A cheaper first pass that triggers frequent fallback might cost more than keeping broader excerpts up front.
My adoption criterion would be a reproducible reduction in total cost or end-to-end time with a stable rate of fully supported answers, including high-consequence exceptions. Context Focus offers a useful way to control what the generator reads. The acceptance evidence is whether the reader still gets everything needed to reach the right conclusion.