headless-intelligent-search（无头智能搜索）

Intelligent Search API Integration

When this skill applies

Use this skill when implementing product search, category browsing, autocomplete, or faceted filtering in a headless VTEX storefront.

• Building a search results page with product listings
• Implementing faceted navigation (category, brand, price, color filters)
• Adding autocomplete suggestions to a search input
• Wiring up search analytics events for Intelligent Search ranking

Do not use this skill for:

• BFF architecture and API routing decisions (use headless-bff-architecture)
• Checkout or cart API integration (use headless-checkout-proxy)
• Caching strategy and TTL configuration (use headless-caching-strategy)

Decision rules

• Call Intelligent Search directly from the frontend — it is the ONE exception to the "everything through BFF" rule. It is fully public and requires no authentication.
• Do NOT proxy Intelligent Search through the BFF unless you have a specific need (e.g., server-side rendering). Proxying adds latency on a high-frequency operation.
• Always use the API's sort parameter and facet paths for filtering and sorting — never re-sort or re-filter results client-side.
• Always include page and locale parameters in every search request.
• When paginating beyond the first page, always reuse the operator and fuzzy values returned by the API from the first page — the API decides these values dynamically based on the search query.
• Always send analytics events to the Intelligent Search Events API — without them, search ranking degrades over time.

Search endpoints overview:

Facet combination rules:

• Same facet type → OR (union): Selecting "Red" and "Blue" for color returns products matching either color
• Different facet types → AND (intersection): Selecting "Red" color and "Nike" brand returns only red Nike products

Hard constraints

Constraint: MUST send analytics events to Intelligent Search Events API

Every headless search implementation MUST send analytics events to the Intelligent Search Events API - Headless. At minimum, send search impression events when results are displayed and click events when a product is selected from search results.

Why this matters

Intelligent Search uses machine learning to rank results based on user behavior. Without analytics events, the search engine has no behavioral data and cannot personalize or optimize results. Over time, search quality degrades compared to stores that send events. Additionally, VTEX Admin search analytics dashboards will show no data.

Detection

If a search implementation renders results from Intelligent Search but has no calls to sp.vtex.com/event-api or the Intelligent Search Events API → STOP immediately. Analytics events must be implemented alongside search.

Correct

// search-analytics.ts — sends events to Intelligent Search Events API
const ACCOUNT_NAME = "mystore";
const EVENTS_URL = `https://sp.vtex.com/event-api/v1/${ACCOUNT_NAME}/event`;

interface SearchEvent {
  type: "search.query" | "search.click" | "search.add_to_cart";
  text: string;
  misspelled: boolean;
  match: number;
  operator: string;
  locale: string;
  agent: string;
  url: string;
  products?: Array<{ productId: string; position: number }>;
}

export async function sendSearchEvent(event: SearchEvent): Promise<void> {
  try {
    await fetch(EVENTS_URL, {
      method: "POST",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify(event),
      keepalive: true, // ensures event is sent even during navigation
    });
  } catch (error) {
    // Analytics failures should not break the UI
    console.warn("Failed to send search event:", error);
  }
}

// Usage: call after rendering search results
function onSearchResultsRendered(query: string, products: Product[]): void {
  sendSearchEvent({
    type: "search.query",
    text: query,
    misspelled: false,
    match: products.length,
    operator: "and",
    locale: "en-US",
    agent: "my-headless-store",
    url: window.location.href,
    products: products.map((p, i) => ({
      productId: p.productId,
      position: i + 1,
    })),
  });
}

Wrong

// Search works but NO analytics events are sent — search ranking degrades
async function searchProducts(query: string): Promise<Product[]> {
  const response = await fetch(
    `https://mystore.vtexcommercestable.com.br/api/io/_v/api/intelligent-search/product_search/?query=${query}&locale=en-US`
  );
  const data = await response.json();
  return data.products;
  // Missing: no call to sendSearchEvent() — Intelligent Search cannot learn
}

Constraint: MUST persist operator and fuzzy values across pagination

When paginating search results beyond the first page, you MUST reuse the operator and fuzzy values returned by the API in the first page response. The first request should send these as null, but subsequent pages must use the values from the initial response.

Why this matters

The Intelligent Search API dynamically determines the best operator and fuzzy values based on the query and available results. Using fixed values (like always operator: "or" or fuzzy: "0") or forgetting to pass them on subsequent pages causes inconsistent results — page 2 may show products that don't match the criteria from page 1, or may miss valid results. This is a common source of poor search experience but is not well-documented by VTEX.

Detection

If pagination implementation does not store and reuse operator/fuzzy from the first page response → STOP immediately. Subsequent pages will return inconsistent results.

Correct

// Properly manages operator/fuzzy across pagination
interface SearchState {
  query: string;
  page: number;
  count: number;
  locale: string;
  facets?: string;
  operator: string | null; // null for first page, then persisted
  fuzzy: string | null;    // null for first page, then persisted
}

async function searchProducts(state: SearchState): Promise<SearchResponse> {
  const { query, page, count, locale, facets = "", operator, fuzzy } = state;

  const params = new URLSearchParams({
    query,
    locale,
    page: String(page),
    count: String(count),
  });

  // Only add operator/fuzzy if this is not the first page
  if (operator !== null) params.set("operator", operator);
  if (fuzzy !== null) params.set("fuzzy", fuzzy);

  const baseUrl = `https://${ACCOUNT}.vtexcommercestable.com.br`;
  const facetPath = facets ? `/${facets}` : "";
  const url = `${baseUrl}/api/io/_v/api/intelligent-search/product_search${facetPath}?${params}`;

  const response = await fetch(url);
  const data = await response.json();

  // Store operator/fuzzy from response for next page
  if (page === 0) {
    state.operator = data.operator;
    state.fuzzy = data.fuzzy;
  }

  return data;
}

// Usage: first page
const state: SearchState = {
  query: "running shoes",
  page: 0,
  count: 24,
  locale: "en-US",
  operator: null, // API will decide
  fuzzy: null,    // API will decide
};

const firstPage = await searchProducts(state);

// Usage: subsequent pages
state.page = 1;
const secondPage = await searchProducts(state); // reuses operator/fuzzy

Wrong

// Using fixed operator/fuzzy values — causes bad results
async function searchProducts(query: string, page: number): Promise<SearchResponse> {
  const params = new URLSearchParams({
    query,
    locale: "en-US",
    page: String(page),
    count: "24",
    operator: "or", // WRONG: fixed value instead of API-provided
    fuzzy: "0",     // WRONG: fixed value instead of API-provided
  });

  const response = await fetch(
    `https://mystore.vtexcommercestable.com.br/api/io/_v/api/intelligent-search/product_search/?${params}`
  );
  return response.json();
}

// Or forgetting operator/fuzzy entirely on subsequent pages
async function searchProductsPage2(query: string): Promise<SearchResponse> {
  const params = new URLSearchParams({
    query,
    locale: "en-US",
    page: "1", // WRONG: no operator/fuzzy from first page
    count: "24",
  });
  // ...
}

Constraint: Do NOT unnecessarily proxy Intelligent Search through BFF

Intelligent Search API requests SHOULD be made directly from the frontend. Do not route search traffic through the BFF unless you have a specific need (e.g., server-side rendering, adding custom business logic).

Why this matters

Intelligent Search is a public API that does not require authentication. Adding a BFF proxy layer introduces an additional network hop, increases latency on every search operation, adds server cost, and prevents the CDN from caching responses efficiently. Search queries are high-frequency operations — even 50ms of added latency impacts conversion.

Detection

If all Intelligent Search calls go through a BFF endpoint instead of directly to VTEX → note this to the developer. It is not a security issue but a performance concern. If there is no justification (like SSR), recommend direct frontend calls.

Correct

// Frontend — calls Intelligent Search directly (no BFF needed)
const VTEX_SEARCH_BASE = `https://${ACCOUNT}.vtexcommercestable.com.br/api/io/_v/api/intelligent-search`;

export async function getAutocomplete(term: string, locale: string): Promise<AutocompleteResponse> {
  const params = new URLSearchParams({ query: term, locale });
  const response = await fetch(`${VTEX_SEARCH_BASE}/autocomplete_suggestions?${params}`);
  return response.json();
}

export async function getTopSearches(locale: string): Promise<TopSearchesResponse> {
  const params = new URLSearchParams({ locale });
  const response = await fetch(`${VTEX_SEARCH_BASE}/top_searches?${params}`);
  return response.json();
}

export async function getFacets(facetPath: string, query: string, locale: string): Promise<FacetsResponse> {
  const params = new URLSearchParams({ query, locale });
  const response = await fetch(`${VTEX_SEARCH_BASE}/facets/${facetPath}?${params}`);
  return response.json();
}

Wrong

// BFF proxy for Intelligent Search — unnecessary overhead
// server/routes/search.ts
router.get("/api/bff/search", async (req, res) => {
  const { query, from, to, locale } = req.query;
  // This just forwards to VTEX with no added value
  const response = await fetch(
    `https://mystore.vtexcommercestable.com.br/api/io/_v/api/intelligent-search/product_search/?query=${query}&from=${from}&to=${to}&locale=${locale}`
  );
  const data = await response.json();
  res.json(data); // Added latency for no benefit
});

Preferred pattern

Data flow for Intelligent Search in a headless storefront:

Frontend (Browser)
    │
    ├── GET /api/io/_v/api/intelligent-search/product_search/...
    │   └── Returns: products, pagination info, operator, fuzzy
    │
    ├── GET /api/io/_v/api/intelligent-search/facets/...
    │   └── Returns: available filters with counts
    │
    ├── GET /api/io/_v/api/intelligent-search/autocomplete_suggestions?query=...
    │   └── Returns: suggested terms + suggested products
    │
    └── POST https://sp.vtex.com/event-api/v1/{account}/event
        └── Sends: search impressions, clicks, add-to-cart events

Typed search API client for all endpoints:

// lib/intelligent-search-client.ts
const ACCOUNT = "mystore";
const ENVIRONMENT = "vtexcommercestable";
const BASE_URL = `https://${ACCOUNT}.${ENVIRONMENT}.com.br/api/io/_v/api/intelligent-search`;

interface ProductSearchParams {
  query?: string;
  page: number;
  count?: number;
  locale: string;
  facets?: string;
  sort?: "price:asc" | "price:desc" | "orders:desc" | "name:asc" | "name:desc" | "release:desc" | "discount:desc";
  hideUnavailableItems?: boolean;
  operator?: string | null;
  fuzzy?: string | null;
  // Optional parameters
  simulationBehavior?: "default" | "skip" | "only1P";
}

interface ProductSearchResponse {
  products: Product[];
  recordsFiltered: number;
  // Pagination info
  pagination: {
    count: number;
    current: { index: number };
    before: Array<{ index: number }>;
    after: Array<{ index: number }>;
    perPage: number;
    next: { index: number } | null;
    previous: { index: number } | null;
    first: { index: number };
    last: { index: number };
  };
  // Search metadata to persist for next pages
  operator: string;
  fuzzy: string;
  // Spelling correction info
  correction?: {
    misspelled: boolean;
    text: string;
    correction: string;
  };
  // Search behavior
  translated: boolean;
  locale: string;
  query: string;
}

export async function productSearch(params: ProductSearchParams): Promise<ProductSearchResponse> {
  const { facets = "", operator, fuzzy, ...queryParams } = params;
  const searchParams = new URLSearchParams();

  if (queryParams.query) searchParams.set("query", queryParams.query);
  searchParams.set("page", String(queryParams.page));
  searchParams.set("locale", queryParams.locale);
  if (queryParams.count) searchParams.set("count", String(queryParams.count));
  if (queryParams.sort) searchParams.set("sort", queryParams.sort);
  if (queryParams.hideUnavailableItems) searchParams.set("hideUnavailableItems", "true");
  if (queryParams.simulationBehavior) searchParams.set("simulationBehavior", queryParams.simulationBehavior);

  // Only add operator/fuzzy if not null (i.e., not first page)
  if (operator !== null && operator !== undefined) searchParams.set("operator", operator);
  if (fuzzy !== null && fuzzy !== undefined) searchParams.set("fuzzy", fuzzy);

  const facetPath = facets ? `/${facets}` : "";
  const url = `${BASE_URL}/product_search${facetPath}?${searchParams}`;

  const response = await fetch(url);
  if (!response.ok) {
    throw new Error(`Search failed: ${response.status}`);
  }
  return response.json();
}

Response structure notes:

• operator and fuzzy: Must be stored from the first page and reused on subsequent pages. Also critical for analytics events.
• pagination: Contains rich pagination metadata including available pages before/after current
• correction.misspelled: Boolean indicating if the search term is misspelled. Must be sent accurately in analytics events.
• recordsFiltered: Total number of results matching the search (across all pages)
• products: Array of products for the current page only

Faceted navigation helper:

// lib/facets.ts
export async function getFacets(
  facetPath: string,
  query: string,
  locale: string
): Promise<FacetsResponse> {
  const params = new URLSearchParams({ query, locale });
  const url = `${BASE_URL}/facets/${facetPath}?${params}`;

  const response = await fetch(url);
  if (!response.ok) {
    throw new Error(`Facets fetch failed: ${response.status}`);
  }
  return response.json();
}

// Build facet path from selected filters
export function buildFacetPath(selectedFilters: Record<string, string[]>): string {
  const parts: string[] = [];
  for (const [key, values] of Object.entries(selectedFilters)) {
    for (const value of values) {
      parts.push(`${key}/${value}`);
    }
  }
  return parts.join("/");
}

Debounced autocomplete for search inputs:

// lib/autocomplete.ts
export function createDebouncedAutocomplete(delayMs: number = 300) {
  let timeoutId: ReturnType<typeof setTimeout> | null = null;

  return function debouncedAutocomplete(
    query: string,
    locale: string,
    callback: (suggestions: AutocompleteResponse) => void
  ): void {
    if (timeoutId) clearTimeout(timeoutId);

    if (query.length < 2) {
      callback({ searches: [] });
      return;
    }

    timeoutId = setTimeout(async () => {
      const params = new URLSearchParams({ query, locale });
      const response = await fetch(`${BASE_URL}/autocomplete_suggestions?${params}`);
      const suggestions = await response.json();
      callback(suggestions);
    }, delayMs);
  };
}

Full search orchestration with analytics:

// search-page.ts — framework-agnostic search orchestration
import { productSearch } from "./lib/intelligent-search-client";
import { getFacets, buildFacetPath } from "./lib/facets";
import { createDebouncedAutocomplete } from "./lib/autocomplete";
import { sendSearchEvent } from "./search-analytics";

interface SearchState {
  query: string;
  page: number;
  count: number;
  locale: string;
  selectedFilters: Record<string, string[]>;
  operator: string | null;
  fuzzy: string | null;
  results: ProductSearchResponse | null;
  facets: FacetsResponse | null;
}

const state: SearchState = {
  query: "",
  page: 0,
  count: 24,
  locale: "en-US",
  selectedFilters: {},
  operator: null, // API decides on first page
  fuzzy: null,    // API decides on first page
  results: null,
  facets: null,
};

async function executeSearch(): Promise<void> {
  const facetPath = buildFacetPath(state.selectedFilters);

  const [searchResults, facetResults] = await Promise.all([
    productSearch({
      query: state.query,
      page: state.page,
      count: state.count,
      locale: state.locale,
      facets: facetPath,
      operator: state.operator,
      fuzzy: state.fuzzy,
    }),
    getFacets(facetPath, state.query, state.locale),
  ]);

  // Store operator/fuzzy from first page for subsequent pages
  if (state.page === 0) {
    state.operator = searchResults.operator;
    state.fuzzy = searchResults.fuzzy;
  }

  state.results = searchResults;
  state.facets = facetResults;

  // Send analytics event after results are rendered
  sendSearchEvent({
    type: "search.query",
    text: state.query,
    misspelled: searchResults.correction?.misspelled ?? false,
    match: searchResults.recordsFiltered,
    operator: searchResults.operator,
    locale: state.locale,
    agent: "my-headless-store",
    url: window.location.href,
    products: searchResults.products.map((p, i) => ({
      productId: p.productId,
      position: state.page * searchResults.products.length + i + 1,
    })),
  });
}

// Handle product click from search results — send click event
function onProductClick(productId: string, position: number): void {
  sendSearchEvent({
    type: "search.click",
    text: state.query,
    misspelled: false,
    match: state.results?.recordsFiltered ?? 0,
    operator: state.results?.operator ?? "and",
    locale: state.locale,
    agent: "my-headless-store",
    url: window.location.href,
    products: [{ productId, position }],
  });
}

// Handle page change
function goToPage(newPage: number): void {
  state.page = newPage;
  executeSearch(); // reuses operator/fuzzy from first page
}

// Handle new search (reset pagination state)
function newSearch(query: string): void {
  state.query = query;
  state.page = 0;
  state.operator = null; // reset for new search
  state.fuzzy = null;    // reset for new search
  executeSearch();
}

Common failure modes

• Not tracking correction.misspelled and operator fields for analytics: The API returns correction.misspelled (boolean indicating if the term is misspelled) and operator fields that must be sent correctly in analytics events. These values are critical for the Intelligent Search machine learning to understand search quality and user behavior. const response = await productSearch({query: "red nike shoes", page: 0, locale: "en-US"}); // Extract values from API response const isMisspelled = response.correction?.misspelled?? false; const operatorUsed = response.operator; // "and" or "or" // CRITICAL: Send these exact values to analytics sendSearchEvent({type: "search.query", text: response.query, misspelled: isMisspelled, // Must match API's correction.misspelled operator: operatorUsed, // Must match API's operator value match: response.recordsFiltered, locale: state.locale, agent: "my-headless-store", url: window.location.href, products: response.products.map((p, i) => ({productId: p.productId, position: i + 1,})),});
• Not sending the locale parameter: Without locale, Intelligent Search may return results in the wrong language or fail to apply locale-specific relevance rules. Multi-language stores will display mixed-language results. Always include locale in every request. // Always include locale in search parameters const params = new URLSearchParams({query: "shoes", locale: "en-US", // Required for correct language processing page: "0", count: "24",});
• Using fixed operator/fuzzy values or forgetting to persist them: Setting fixed values like operator: "or" or fuzzy: "0" for all requests, or not passing them at all on subsequent pages. The API dynamically determines these values on the first page — you must store and reuse them for pagination. // Store operator/fuzzy from first page, reuse on subsequent pages let searchState = {operator: null, fuzzy: null}; const firstPage = await productSearch({query: "shoes", page: 0, count: 24, locale: "en-US"}); searchState.operator = firstPage.operator; // Store from API response searchState.fuzzy = firstPage.fuzzy; // Page 2 reuses these values const secondPage = await productSearch({query: "shoes", page: 1, count: 24, locale: "en-US", operator: searchState.operator, // Reuse from first page fuzzy: searchState.fuzzy,});
• Assuming /product_search returns facets: The /product_search endpoint returns only products and pagination info. To get available filters, you must call the separate /facets endpoint. Make parallel requests if you need both. // Fetch products and facets in parallel const [products, filters] = await Promise.all([productSearch({query: "shoes", page: 0, locale: "en-US"}), getFacets("", "shoes", "en-US"), // Separate call for filters]);
• Not using hideUnavailableItems when appropriate: By default, the API may return out-of-stock products. For most storefronts, set hideUnavailableItems: true to filter them out at the API level rather than client-side. const response = await productSearch({query: "shoes", page: 0, locale: "en-US", hideUnavailableItems: true, // Filter out-of-stock products});
• Rebuilding search ranking logic client-side: Fetching results and then re-sorting or re-filtering them in the frontend discards Intelligent Search's ranking intelligence. Client-side filtering only works on the current page, not the full catalog. Use the API's sort parameter and facet paths. // Use API-level sorting — don't re-sort in the frontend const results = await productSearch({query: "shirt", sort: "price:asc", // API handles sorting across entire result set locale: "en-US", from: 0, to: 23, facets: "category-1/clothing/color/red", // API handles filtering across entire catalog});

Review checklist

• Is Intelligent Search called directly from the frontend (not unnecessarily routed through BFF)?
• Does every search request include page and locale parameters?
• Are operator and fuzzy values stored from the first page and reused on subsequent pages?
• Are correction.misspelled and operator from the API response sent correctly in analytics events?
• Are analytics events sent to sp.vtex.com/event-api after search results render?
• Are click events sent when a user selects a product from search results?
• Is sorting done via the API's sort parameter rather than client-side re-sorting?
• Is filtering done via facet paths rather than client-side filtering?
• Is autocomplete debounced to avoid excessive API calls?
• Are filters fetched from the /facets endpoint (not assumed to come from /product_search)?
• Is hideUnavailableItems set to true for storefronts that should hide out-of-stock products?

Reference

• Headless catalog and search — Overview of catalog browsing and search in headless stores
• Intelligent Search API reference — Complete API reference for all search endpoints
• Intelligent Search Events API - Headless — Events API for sending analytics from headless implementations
• Intelligent Search overview — General overview of Intelligent Search capabilities
• Search configuration — How to configure searchable specifications, facet ordering, and other search settings
• Autocomplete — How autocomplete suggestions work in Intelligent Search