Skip to content

An AI-powered search engine with a generative UI

fix(search): project results onto their declared shape for the model (#1032)

Yoshiki Miura30e809b789d1
diff --git a/lib/tools/__tests__/search-to-model-output.test.ts b/lib/tools/__tests__/search-to-model-output.test.ts
index 82fc1f5e..9e84c840 100644
--- a/lib/tools/__tests__/search-to-model-output.test.ts
+++ b/lib/tools/__tests__/search-to-model-output.test.ts
@@ -69,19 +69,69 @@ describe('search tool toModelOutput', () => {
expect(value).not.toHaveProperty('toolCallId')
})
+ it('keeps only model-facing fields on each result', async () => {
+ const result = {
+ title: 'A',
+ url: 'https://a.test',
+ content: 'alpha',
+ label: 'S1',
+ id: 'provider-id',
+ score: 0.37,
+ raw_content: null,
+ images: []
+ }
+ const value = await getModelValue({
+ ...fullOutput,
+ results: [result]
+ })
+
+ expect(value.results).toEqual([
+ {
+ title: result.title,
+ url: result.url,
+ content: result.content,
+ label: result.label
+ }
+ ])
+ expect(value.images).toBe(fullOutput.images)
+ })
+
+ it('does not add a label to an unlabelled result', async () => {
+ const value = await getModelValue({
+ ...fullOutput,
+ results: [{ title: 'A', url: 'https://a.test', content: 'alpha' }]
+ })
+ const projectedResult = (value.results as Array<Record<string, unknown>>)[0]
+
+ expect(projectedResult).not.toHaveProperty('label')
+ })
+
it('truncates long result content and appends a marker inside the bound', async () => {
const content = 'a'.repeat(SEARCH_MODEL_CONTENT_MAX_CHARACTERS + 10)
const value = await getModelValue({
...fullOutput,
- results: [{ ...fullOutput.results[0], content }]
+ results: [
+ {
+ ...fullOutput.results[0],
+ content,
+ id: 'provider-id',
+ score: 0.37,
+ raw_content: null,
+ images: []
+ }
+ ]
})
- const projectedContent = (value.results as Array<{ content: string }>)[0]
- .content
+ const projectedResult = (value.results as Array<Record<string, unknown>>)[0]
+ const projectedContent = projectedResult.content as string
expect(projectedContent).toBe(
`${content.slice(0, SEARCH_MODEL_CONTENT_MAX_CHARACTERS - 1)}…`
)
expect(projectedContent).toHaveLength(SEARCH_MODEL_CONTENT_MAX_CHARACTERS)
+ expect(projectedResult).not.toHaveProperty('id')
+ expect(projectedResult).not.toHaveProperty('score')
+ expect(projectedResult).not.toHaveProperty('raw_content')
+ expect(projectedResult).not.toHaveProperty('images')
})
it('does not split a surrogate pair at the boundary', async () => {
diff --git a/lib/tools/search.ts b/lib/tools/search.ts
index 6bf111fa..f58a60b5 100644
--- a/lib/tools/search.ts
+++ b/lib/tools/search.ts
@@ -26,6 +26,10 @@ import {
// ellipsis is counted inside the bound, as it is for historical messages.
export const SEARCH_MODEL_CONTENT_MAX_CHARACTERS = 600
+// The declared SearchResultItem shape. Everything else a provider returns is
+// kept out of the model-facing projection.
+const MODEL_FACING_RESULT_FIELDS = ['url', 'title', 'content', 'label'] as const
+
function getOptimizedSearchProviderType(): SearchProviderType {
return (process.env.SEARCH_API as SearchProviderType) || DEFAULT_PROVIDER
}
@@ -254,18 +258,14 @@ export function createSearchTool(
...(fallback ? { fallback } : {})
}
},
- // Trim the model-facing tool result: citationMap fully duplicates
- // `results` (dropped defensively for older persisted output), state is a
- // streaming marker, and provider/fallback are trace diagnostics.
- // toolCallId is dropped too: citations address a result's own `label`, so
- // the opaque id no longer belongs in the model's view. Result content is
- // capped, but every result is kept: dropping results would decide relevance
- // from provider order, and providers do not all order by relevance (the
- // firecrawl adapter appends news after web). `execute` still yields the
- // untruncated content, so the UI and the persisted citation targets are
- // unaffected. images MUST stay -
- // getImageSpecPrompt instructs the model to embed URLs verbatim from that
- // array. Labels are assigned in `execute` and only passed through here.
+ // Trim the model-facing tool result: the deleted keys are duplicates or
+ // diagnostics the model cannot cite. Results are narrowed by allowlist
+ // rather than by a delete list, because a provider's own result id would
+ // otherwise sit beside `label` as a competing citation target that
+ // resolves to nothing. Content is capped but no result is dropped, since
+ // provider order is not relevance order. `execute` still yields the full
+ // result, so the UI and persisted citations are unaffected. images MUST
+ // stay: getImageSpecPrompt has the model copy URLs verbatim from it.
toModelOutput: ({ output }) => {
if (!output || typeof output !== 'object') {
return { type: 'json', value: (output ?? null) as JSONValue }
@@ -284,8 +284,12 @@ export function createSearchTool(
return result
}
- const projectedResult = {
- ...(result as Record<string, unknown>)
+ const sourceResult = result as Record<string, unknown>
+ const projectedResult: Record<string, unknown> = {}
+ for (const field of MODEL_FACING_RESULT_FIELDS) {
+ if (Object.hasOwn(sourceResult, field)) {
+ projectedResult[field] = sourceResult[field]
+ }
}
if (
typeof projectedResult.content === 'string' &&