search-integration

Search and Knowledge Retrieval Integration

R — Original Text (Reading)

Search strategy patterns emerging from cross-vendor system prompts: Claude Web's "confidence is not a reason to skip searching," Perplexity's "follow-up questions must trigger a new search," Gemini's multi-query strategy (at least one question-style + one keyword-style), Notion AI's "searching is cheap, search by default on every first interaction." Le Chat mandates searching for all contemporary public figures, while NotebookLM does not search at all—pure source document citation. The core point of divergence lies in the threshold for searching: from "don't search unless you have to" to "search by default every time."

I — Methodological Skeleton (Interpretation)

1. Search-first strategy (search_first) — For time-sensitive information, factual claims, and contemporary people/events, searching is the default action rather than an optional one
2. Multi-query combination — A single search should issue at least two different forms of queries: natural-language question style + keyword style, covering different indexing modes
3. Source priority hierarchy — Enterprise data > licensed corpora > public web search > social media, with the hierarchy defined by scenario
4. Follow-up re-search principle — Do not assume the previous results are still valid for a follow-up; re-retrieve on every substantive question
5. Domain-specific search rules — Finance = single-entity focus, local = geocoding, travel = transportation + hotels, sports = full context
6. Search cost awareness — "Searching is cheap, safe, and fast, and users are willing to wait" (Notion AI), lowering the threshold for searching
7. No-search exception — Pure source-document scenarios (NotebookLM) replace searching with sentence-by-sentence citations to ensure faithfulness

A1 — Case Analysis (Past Application)

Case: Gemini's mandatory multi-query strategy

• Problem: A single query cannot cover the different expression dimensions of user intent, causing search results to miss key information
• Use of the design pattern: The Gemini 3.1 Pro system prompt requires at least two queries per search: one natural-language question-style query (capturing semantics) plus one keyword-style query (capturing exact matches), and all queries must use the user's original language
• Conclusion: The multi-query strategy upgrades information retrieval from "guess one best query" to "multi-angle coverage," significantly reducing the information omission rate

Case: Perplexity's follow-up re-search principle

• Problem: When users ask follow-up questions, the system tends to reuse previous search results to save time and tokens, but the information may already be outdated
• Use of the design pattern: The Perplexity system prompt explicitly states: "Always search again on follow-up questions, rather than assuming the previous results are still sufficient"
• Conclusion: This strategy sacrifices efficiency for accuracy, and its effect is especially significant in highly time-sensitive scenarios such as news and finance

Case: NotebookLM's anti-search pattern

• Problem: General search may introduce external information that pollutes faithful interpretation of source documents
• Use of the design pattern: NotebookLM does not search at all, strictly based on the source documents uploaded by the user, combined with a sentence-by-sentence citation mechanism to ensure every claim is traceable
• Conclusion: In scenarios requiring high fidelity (academic analysis, legal documents), "not searching" is instead the correct strategy

A2 — Trigger Scenario (Future Trigger) ★

In what situations do users need this?

1. Designing system prompts for AI assistants that need internet access, requiring definition of "when to search, what to search, and how many to search"
2. Building an enterprise knowledge base question-answering system, requiring definition of the priority relationship between internal corpora and external search
3. Designing search strategies for vertical-domain AI (finance, healthcare, law), requiring domain-specific rules
4. Optimizing the search trigger rate of existing AI products—users report "outdated information" or "answers lack the latest data"
5. Designing research-oriented AI products (such as Deep Research), requiring multi-round search and result integration strategies

Language Signals

• "The information in the AI's answer is outdated"
• "Needs to cite the latest data/news"
• "Search first, then answer; don't rely on memory"
• "Internal enterprise knowledge takes priority over web search"
• "Every follow-up question must be looked up again"

Distinction from adjacent skills

• Difference from context-management : context-management manages the compression and loading of existing context; this Skill manages the timing and strategy for acquiring external information
• Difference from conversation-flow : conversation-flow manages dialogue routing and clarification strategies; this Skill focuses on search decisions (whether to search, how many to search, where to search)

E — Executable Steps (Execution)

1. Define the search trigger rule matrix — Completion criteria: Establish a two-dimensional matrix based on content type (timeliness/factuality/person/viewpoint) and timeliness requirements (real-time/recent/historical), clarifying the search strategy for each combination (mandatory search/suggested search/optional search/prohibited search)

2. Design multi-query combination templates — Completion criteria: Define at least two query forms (question-style + keyword-style) for each search trigger point, including language-preservation rules (use the user's original language) and query expansion strategies

3. Establish source priority tiers — Completion criteria: Define at least three tiers of source priority (e.g., enterprise corpus > authorized databases > public web pages), including adjudication rules and fallback strategies for cross-source conflicts

4. Write follow-up search strategies — Completion criteria: Clearly specify search behavior in follow-up scenarios—at minimum distinguish between "information-supplementing follow-ups" (search again) and "logic-clarifying follow-ups" (do not search), including time-window judgment (must search again after more than N minutes)

5. Add domain-specific search parameters — Completion criteria: Define dedicated search parameter sets for vertical domains such as finance (single-entity focus + time range), local (geocoding + distance radius), travel (joint queries for transportation + accommodation + weather), and sports (full event context)

B — Boundary ★

Do not use in the following situations

• Pure math/logic reasoning tasks—searching will not be more accurate than model reasoning
• Creative writing/literary creation—external search may disrupt creative consistency
• Closed-document Q&A with full context already available (NotebookLM mode)—search introduces noise
• Historical/philosophical discussions with extremely low real-time requirements—the model's internalized knowledge is sufficient

Common failure modes

• Over-searching: searching on every reply even when the user is just chatting, increasing latency and cost with no information gain
• Single-query reliance: using only one query form, missing information from different index dimensions
• Inverted source priority: public web information overrides authoritative internal enterprise data, causing answers to deviate from business facts
• Not re-searching on follow-up: reusing outdated results to answer follow-up questions, producing hallucinations in news/financial scenarios
• Ignoring language consistency: searching in English and then translating back to Chinese, losing query precision and cultural context