Large model application network search: reshaping interaction and decision-making in the intelligent era

In the wave of artificial intelligence technology, the performance competition of large models continues to heat up. DeepSeek-R1 has swept the world with its powerful reasoning ability, and the open source of Tongyi Qianwen QwQ has injected new vitality into the industry. However, a key problem has surfaced: neither DeepSeek-R1 (when deployed by itself) nor Alibaba Cloud's latest QwQ (currently using API calls) supports the [online search] capability. This means that the knowledge boundaries of these models are strictly limited to local training data or closed knowledge bases , and cannot capture the massive dynamic information of the Internet in real time.

• Core point:
  Currently, there are two significant watersheds in the application of large models: those with online search capabilities and those without such capabilities. The latter has obvious shortcomings in generation quality, timeliness, and user trust. Statistics show that access to online search can increase the accuracy of model output by more than 50% and user satisfaction by more than 30%.

• Trend driving:
  Developers who are deeply engaged in enterprise-level AI scenarios have gradually reached a consensus: "AI without Internet access is like a tree without roots." Alibaba Cloud's cloud-native API gateway (AI gateway) is redefining the standards for intelligent services by deeply integrating Internet search capabilities.

• Dynamic data acquisition: Break through the data timeliness limit during model training and capture information from trusted sources such as web pages, databases, and APIs in real time
• Scenario examples: Real-time access to financial news in the financial industry and dynamic query of the latest clinical guidelines in the medical scenario
• Technical implementation: The cloud-native API gateway provides multi-engine network search capabilities and completes cross-source fusion within 1 second.

2. Complex Problem Solver: From “Answering Questions” to “Solving Problems”

• Multi-round dialogue enhancement: Complete missing information in the process (such as order number, express delivery status) through search.
• Big data association reasoning: Analyze implicit relationships in search results and output structured solutions.
• Scenario example: The customer service system automatically links the user's orders and logistics information from the past three months to resolve complaints.

3. Intelligent cost optimization: semantic caching and dynamic routing

• Duplicate request interception: By configuring the cache service, common questions can be directly responded to through the cache, reducing the API call cost by 25%.

• Multi-model intelligent scheduling: automatically matches basic model/professional large model/search enhancement mode according to query complexity.

Advantage 1: Real-time and dynamic

• Do not rely on local cache and directly connect to the Internet to obtain the latest data (such as emergencies and industry news).
• Case comparison: Traditional engine searches for "South Korean chip export data 2024Q2" may rely on old statistics, while AI networked searches can capture the latest announcements from the South Korean Ministry of Industry in real time.

• Handle multi-condition combination and implicit logic query, for example: "List the provinces in China that provide new energy battery R&D support policies, and analyze the policy effective time and subsidy amount."
• Technical support: The semantic understanding capability of the large model is combined with the rule engine to achieve accurate analysis.

• Customize information priority and presentation based on user role (analyst, customer service, executive).
• Case: Provide structured data (such as user feedback hot issues + solutions) to customer service robots to improve response speed and accuracy.

Challenge 1: Reliability and real-time performance of data sources

• Problem: The quality of Internet data varies greatly, and real-time crawling faces performance bottlenecks.
• Solution:

• Intelligent filtering and verification: Filter valid information through semantic analysis and credibility scoring (such as source authority).
• Incremental update mechanism: Focus on monitoring updates in key areas (such as finance and healthcare) to reduce network-wide scanning overhead.

• Problem: The external data searched may involve sensitive political or violent information.
• Solution:

• Green Network interception mechanism: By configuring the Green Network security service, user input and search results can be uniformly filtered for content security.
• Consumer authorization system: Only authorized users have API access qualifications, and access rights can be planned and configured in a fine-grained manner.

• Problem: Real-time online search may trigger high costs for downloading large amounts of data and reasoning about large models.
• Solution:

• Summary input: By default, the summary information of the search results is used to fill the prompt word to prevent the context window from being exhausted quickly.

• Cache optimization: cache high-frequency query results to reduce repeated reasoning and network requests.

1. Log in to the Cloud Native API Gateway console ^{[ 1]} .
2. In the left navigation bar, select API and select Region in the top menu bar.
3. In the AI ​​API list, click the target API to go to the target API details page.
4. Select the Policies & Plugins tab and enable Network Search .

• Number of results returned: 1-10, the maximum value is 10, that is, a maximum of 10 results will be returned
• Timeout: 3000ms by default
• Query time range: within 1 day, within 1 week, within 1 month, within 1 year, unlimited
• Industry (optional): Finance, Law, Medical, Internet, Taxation, Provincial News, Central News

• Default language: Chinese, English
• Output citation source:

• Effect of “No”:

• The effect of "yes"

• Content Type:

• Summary (default): Returns only summary information of the search terms, which is generally sufficient for model acquisition.
• Main text: Returns the main text of the search entry, which contains a large amount of information but is detailed. It is suitable for scenarios where detailed information is required.

• Reference format: %s is the rendering placeholder for the reference entry. The display format of the reference entry can be modified as required.

Trend 1: Deep integration with real-time interactive technology

• WebSocket+AI: Embed AI networked search capabilities into real-time dialogue systems (such as customer service and virtual assistants) to achieve “conversation, search, and feedback”.
• Case: Integrate with games to provide players with cross-platform strategies and the latest event updates.

• Enterprise-level self-built search service access: AI networked search will provide enterprises with the ability to integrate their own search services, helping them to quickly build intelligent products using their own data.
• Case: Banks use AI network search to build compliance risk early warning systems and dynamically monitor changes in regulatory policies.

• Multi-party collaboration: Cooperate with vertical data platforms and developer communities to create standardized and traceable search services.
• Open source and openness: lower the threshold for AI network search technology and promote its application by small and medium-sized enterprises.