News Search API Integration: Separate Retrieval Context From Display Text

A news search API response should not become a customer-facing response simply because the fields are available. An application may retrieve article text for analysis, select passages for a model and show a much smaller result to the user. Give each stage its own data structure and output rules.

The New York Times filed a complaint against Brave Software on November 2, 2026, challenging uses of its journalism across Brave’s products. One API example in the complaint alleges that an LLM Context response reproduced 1,291 words from a 1,295-word article. These are the plaintiff’s allegations, not a court finding or a measurement of every Brave response.

For developers, the practical question is what their own system retrieves, retains and displays. The workflow below is an application-design proposal, not a conclusion about the litigation. Set the permitted uses with the people responsible for your data agreements and product policy before encoding them as software rules.

Keep three different records

Separate the provider response, the model’s evidence bundle and the result returned to the customer. Passing the complete provider object through every layer makes it difficult to see where article text leaves the retrieval system. Instead, select fields explicitly at each boundary.

This table describes proposed application records, not Webz.io response schemas. A field in the retrieval record should enter the display record only through an intentional transformation. Keep diagnostic information and unused source text out of customer payloads, including hidden fields that a browser or export function could expose.

Write an output policy that a renderer can enforce

For each intended use, record the audience, permitted output form, source set, attribution requirements and retention conditions approved for that use. Identify who owns the decision and where the supporting agreement or review is recorded. If a decision is unresolved, route that case to the owner rather than silently choosing permissive defaults.

For example, a fictional test policy might allow an internal analyst card to contain one short approved excerpt and a source link, while a public alert contains only an independently written event description and the link. Those are invented product settings, not legal limits or claims about a Webz.io subscription.

Give those settings a version and apply them outside the model prompt. The model can produce a candidate answer, but the application should decide which fields may be serialized, which excerpt may be included and which source reference must accompany it. A prompt instruction alone is not a release check.

Do not confuse passage relevance with display permission

Webz.io’s News Search API accepts natural-language queries and returns matching article information with relevant content chunks, relevance information and credit usage. It supports filters including source, publication date, language and country. These capabilities help developers select evidence for an AI application.

A relevant passage answers a retrieval question: does this text help with the user’s request? Your application’s approved output policy answers a different question: what should this user receive in this context? Apply that policy to the selected passage and the generated answer rather than assuming a relevance score supplies the decision.

Use the existing discussion of news API source legitimacy for the wider sourcing questions. Then make the implementation specific: identify the audience of this response, the policy version applied and the exact transformation from evidence to display. Keep these decisions attached to the source reference.

Check repeated requests and every delivery channel

Test more than the first answer. A user might ask for the next section, request a longer explanation or repeat a query with slightly different wording. Review the combined text returned from the same source across those interactions, using the aggregation scope defined in your policy.

Keep article identity separate from a single search-result position. The same article can appear for multiple queries or through different returned passages. Use stable provider identifiers where available and a documented URL-normalization rule. Preserve the original references so an engineer can inspect incorrect merges or missed matches.

Check exports, email digests, saved reports and webhook payloads alongside the chat interface. A compact screen can coexist with an export that includes the entire retrieval object. Apply the same approved audience and output rules at each delivery boundary rather than relying on what the main page happens to render.

Validate the final answer and preserve its evidence

Inspect generated output after generation. Check for unsupported claims, missing source links and text that exceeds the approved excerpt behavior. If the response fails, regenerate within the permitted form or return a controlled failure with the relevant source link, according to your product’s policy.

Where you compare output with source text, document what the comparison catches. Exact phrase matching can detect direct reuse but does not evaluate every paraphrase. A similarity score is another engineering signal, not an automatic determination of permission. Assign review cases that the validation rules cannot resolve.

Record the query, article references, selected evidence identifiers, output-policy version and validation result. Retain source text only under the applicable retention rules. This lets a reviewer reconstruct why a result was delivered without making every application log another repository of full articles.

Build regression tests with material you control

Create a fictional article whose text you own and give it a stable test identifier. Exercise short answers, requests for full text, successive-section requests and export downloads. Use the same article under two artificial policy profiles so the tests verify audience-dependent output rather than merely checking response length.

Include a missing-policy case, a missing-source-link case and a result returned under two different URLs. Confirm that the renderer rejects unresolved policy decisions, preserves attribution and applies the documented article-grouping rule. Record any exception with an owner and an expiry date.

A useful release record reads: “Fixture A; internal policy P-4; excerpt allowed; attribution present; repeated-request check passed; export check failed; release blocked.” Fix the failed delivery path, rerun the affected tests and retain the result with the policy version used for the release.