You.com Adds One API Flag That Boosts Search Accuracy by 20%
You.com's Knowledge parameter injects licensed real-time data into web search results, lifting benchmark accuracy from 64.5% to 84.2% with 140ms extra latency.
- You.com launched a Knowledge parameter that returns licensed real-time data alongside web search results.
- Accuracy on the VerticalRTK benchmark jumps from 64.5% to 84.2% with only 140ms added latency.
- Enabled with a single flag:
"knowledge": "core"on the existing Web Search API endpoint. - No extra charge over the standard $5 CPM Web Search pricing.
- Covers seven categories: stock prices, crypto, weather, sports, company data, politics, and macroeconomics.
- Runnable examples and benchmark harness at ydc-knowledge-samples.
You.com puts licensed data behind one Search API flag
You.com has added a "knowledge": "core" parameter to its Web Search API. When enabled, a search request can return structured data from licensed providers, including S&P Global, AccuWeather, CoinMarketCap, The Economist and the U.S. Treasury, alongside crawled web results.
Agents often struggle with questions that require an exact, current value absent from ordinary web pages. Equity quotes come from market feeds, prediction-market odds come from order books, and federal contract obligations live in government databases. Search results may mention those figures without preserving the labels, dates and fields needed to identify the correct answer.
Why this benchmark stalls near 65%
You.com evaluated 155 questions from the public VerticalRTK benchmark, which focuses on live and vertical-specific data. Four search APIs from four companies scored between 63.9% and 65.2%. The slowest endpoint took 4.4 seconds and gained 0.7 percentage points over the fastest.
| Configuration | Accuracy | Reported latency | Price |
|---|---|---|---|
| Web Search API | 64.5% | 0.60 seconds | $5 per 1,000 searches |
Web Search API with knowledge: core |
84.2% | 0.74 seconds | $5 per 1,000 searches |
The added lookup improved accuracy by 19.7 percentage points with 140 milliseconds of reported latency. These figures come from You.com’s own test on a benchmark designed around live and specialized facts. General web-search quality requires a broader evaluation.
A NASA query exposes the retrieval gap
When asked how much NASA recorded in federal contract outlays for fiscal year 2025, the tested search APIs returned values ranging from $19.8 billion to $28.8 billion. According to You.com, each number appeared on a genuine page but represented a different line item. The closest answer missed the target by $200 million.
Crawled pages can contain several related spending figures, while snippets omit the database fields that distinguish obligations, outlays, awards and fiscal periods. For the NASA query, the Knowledge response supplies a Treasury record labeled with its field name, fiscal year and provider attribution.
One flag starts both retrieval paths
Setting "knowledge": "core" runs licensed-data lookups in parallel with the web-index search. The request format otherwise remains unchanged:
curl -X POST https://ydc-index.io/v1/search \
-H "X-API-Key: $YDC_KEY" \
-H "Content-Type: application/json" \
-d '{
"query": "What is Nvidia trading at right now?",
"count": 10,
"knowledge": "core"
}'A licensed record appears under results.knowledge when a match clears You.com’s internal confidence threshold. The response includes:
- A description of the matched record
- An
as_ofdate indicating its freshness - Attribution naming the data provider
- Source-specific labels that clarify the value and reporting period
Applications should treat results.knowledge as optional because a query may fail to match a supported source or clear the threshold. The launch material does not specify whether developers can tune that threshold.
Coverage at launch
Launch coverage spans seven areas: prices, rates, weather, sports, company information, politics and macroeconomics. The published examples include five queries across four data categories.
| Example query | Provider | Structured result |
|---|---|---|
| Nvidia’s current price | Xignite and S&P Global | Equity quote |
| Bitcoin’s current price | CoinMarketCap | Cryptocurrency quote |
| Local weather forecast | AccuWeather | Forecast data |
| Big Mac Index | The Economist | Economic indicator |
| Federal contract outlays | U.S. Treasury | Government spending record |
The sample repository includes runnable versions of all five queries and the benchmark harness. Running make bench executes the same questions with and without Knowledge, grades both result sets and compares their performance.
The router moves into the API
An agent that accepts arbitrary questions cannot predict whether its next request will require a stock quote, weather forecast, economic indicator or government record. A custom implementation would require separate commercial agreements, client integrations, normalized schemas and routing logic for each provider.
The Knowledge parameter moves source selection and normalization behind an existing search request while preserving provider attribution. Coverage remains bounded by supported feeds and confidence matching, so developers still need to handle missing records, inspect as_of dates, define web-search fallbacks and review the applicable storage or redistribution terms.
Where the flag fits
| Workload | Recommended use | Reason |
|---|---|---|
| General-purpose agent with unpredictable queries | Enable by default and monitor matches | Queries may require several licensed-data categories. |
| Article summarization or document retrieval | Leave disabled | Structured lookups are unlikely to return useful records. |
| Latency-sensitive application | Enable selectively | You.com’s benchmark reports an additional 140 milliseconds. |
| Application with strict freshness limits | Validate as_of |
Update schedules depend on the underlying provider. |
The NASA record illustrates static structured data: fiscal year 2025 closed on September 30, 2025. Other providers may supply frequently updated or streaming values. Applications that depend on recency should enforce their own maximum age using the returned as_of field.
Request fields and response details are available in the API reference.