wps-deep-search

Contract

Input Contract

Type: natural language query or structured search request
Required fields:
- query: user question or goal
Optional fields:
- context: conversation context
- filters: time range, tags, project scope, source preference, etc.
- max_results: default 10, range 1-20
- search_depth: quick | standard | deep(default standard）
- source_mode: notes_only | web_only | hybrid(default hybrid）
Validation rules:
- query length >= 2
- When the question contains "latest/today/recent", etc., the web timeliness check must be enabled

Output Contract

Type: structured retrieval results + readable conclusions
Guaranteed fields:
- results: result list (sorted by overall relevance)
- search_dimensions: retrieval dimensions actually executed
- evidence: evidence list (note evidence + external evidence)
- confidence: confidence level for each key conclusion (High/Medium/Low or 0-100)
- conflicts: conflicting information or unresolved points
Constraints:
- Conclusions must be traceable to evidence sources
- Distinguish between "confirmed" and "to be verified"
- Clearly state which conclusions come from notes and which come from external web sources

Error Handling

Invalid query → return optimization suggestions, do not fail directly
No matching results → automatically relax constraints once and retry; if there are still no results, return "insufficient evidence"
Tool failure → return available partial results + unfinished items
Insufficient authoritative sources → degrade output, do not give definitive conclusions

Core principles

Deep search is not single-round keyword matching, but a "heuristic funnel + evidence convergence".
Unless the user explicitly requests "search notes only", use hybridby default (notes + authoritative web sources).
External sources must be traceable,web_search summaries can only serve as clues,web_fetch only the main text can serve as evidence.
It is strictly prohibited to fabricate sources, fabricate citations, or treat a single low-authority source as a definitive conclusion.
This version enforces serial execution with a single agent and prohibits calling sub_agent。
This Skill is a read-only retrieval process and does not call note-writing tools.

Relationship with wps-note

This Skill is wps-note 's deep retrieval scenario encapsulation layer:

Reuses underlying capabilities: search, locate, read, tag aggregation
Adds intelligent capabilities: heuristic expansion, evidence fusion, confidence evaluation
Adds external capabilities: authoritative web source retrieval and cross-validation

Tool list and invocation methods

Note-side tools (read-only)

Tool name	Purpose
`search_notes`	Note-level recall (title/body/tag dimensions)
`search_note_content`	Locate key paragraphs within a single note
`get_note_info`	Retrieve metadata (title, tags, update time)
`find_tags`	Tag expansion and synonym aggregation
`get_note_outline`	Retrieve structure and locate reading entry points
`read_section`	Chapter-level close reading
`read_blocks`	Precise re-reading of evidence blocks
`read_note` / `read_note_content`	Full-text reading (host tool name compatible)

Network-side tools (read-only)

Tool name	Purpose
`web_search`	Single-query retrieval of authoritative candidates
`batch_web_search`	Multi-query parallel recall (recommended)
`web_fetch`	Fetch full text to form citable evidence

Disabled tools

sub_agent(Prohibited for use by this Skill)

Authoritative source tiering (mandatory)

Tier	Definition	Role in conclusions
Tier 1 (primary official)	Government/regulatory/standards organizations/company official disclosures/original institutional releases	Primary anchor for conclusions
Tier 2 (secondary mainstream)	Mainstream media, leading research institutions, sources with traceable authors and publication times	Cross-validation
Tier 3 (tertiary supplementary)	Communities, self-media, secondary reposts, opinion pieces	Clues and background, not sufficient on their own for conclusions

Rules:

Key conclusions prioritize Tier 1; when there is no Tier 1, there must be at least 2 independent Tier 2 sources.
With only Tier 3, output only "exploratory judgment/insufficient evidence".

Heuristic Funnel

Layer 1: Intent Clarification

Goal: turn the user's question into a retrievable task.

Identify task type: fact-checking/relationship mining/trend judgment/decision recommendation
Extract structured elements:实体、时间窗、约束、输出目标
Set source_mode:notes_only | web_only | hybrid
Clarify obvious ambiguity at most once; otherwise proceed based on conservative assumptions and state the assumptions

Layer 2: Candidate Expansion

Goal: fully recall first, then converge with high quality.

Synonym expansion: alternative names, abbreviations/full names, Chinese and English terms
Reverse-term expansion: search positive and negative signals together to reduce one-sided bias
Time anchor expansion: convert relative time to absolute time, and add a comparison window
Entity relationship expansion: upstream and downstream, competitors, regulators, partners, persons in charge
Context expansion: prioritize project names/internal terminology from the user's notes

Layer 3: Evidence Convergence

Goal: make every conclusion auditable.

Extract evidence units: claim, evidence snippet, source, time, scope of applicability
Independent cross-validation: key claims require corroboration from multiple sources
Preserve conflicts: mark points of dispute, do not force consolidation
Output confidence: high/medium/low, and explain the reasons

Workflow (serial)

Step 0: Mode selection

User explicitly says "only look at notes" → notes_only
User emphasizes "latest/external developments/authoritative statements" → hybrid(or web_only）
When unspecified, default to hybrid

Step 1: Break the question down into claims

Break the question into verifiable assertions (claims); each claim must contain at least:

subject
predicate
time_scope

Step 2: Plan retrieval and budget

Each claim defaults to 2 rounds of web search
Each claim defaults to fetching at most 6 pages
Early stopping conditions:
- Already have 1 Tier 1 + 1 independent Tier 2
- Or already have 2 independent Tier 2 sources that are consistent

Step 3: Note-side retrieval (serial)

search_notes / find_tags Recall candidate notes
search_note_content Locate key passages
read_section / read_blocks Extract evidence

Step 4: Web-side retrieval (serial)

batch_web_search Perform multi-query expansion recall
web_fetch Fetch full text and filter out abstract-only results
Keep candidates in the order Tier 1 → Tier 2 → Tier 3

Step 5: Evidence fusion and scoring

Suggested composite score (0-100):

Intent match (30)
Evidence strength (25)
Source authority (20)
Timeliness match (15)
Cross-source consistency (10)

Confidence suggestions:

High: >= 80 and meets authoritative source conditions
Medium：60-79
Low: < 60 or major conflicts exist

Step 6: Result presentation (source transparency)

Must include:

Query topic and retrieval time
Conclusion summary
Claim-level judgment (supported / partially_supported / conflicting / insufficient_evidence)
Note evidence (note_id / location / summary)
External evidence (Tier, source name, URL, publication time, fetch time, role of evidence)
Conflicts and unresolved issues

Common orchestration patterns

Pattern 1: Fact-checking (default hybrid)

Applicable to "Is this true / what is the latest official position":

claim 拆解 → 笔记检索 → 网络检索 → 证据收敛 → 置信度输出

Pattern 2: Cross-note association + external verification

Applicable to "How does this topic relate to my past notes, and what do external sources say":

笔记召回 → 主题扩展 → 网络权威检索 → 一致性分析

Pattern 3: Trend assessment

Applicable to "Recent trends / risk changes":

时间锚点扩展 → 多时间窗检索 → 证据对比 → 趋势结论

Examples

Example 1: Fact-checking

Input：深度搜索：某 AI 公司最近是否发布了新的企业定价政策？

Execution points：

Claim decomposition: whether released / release time / scope involved
Serially execute official source + mainstream media retrieval
Look back at historical pricing discussions in the user's notes
Output claim-level evidence, confidence, and conflict points

Example 2: Association mining

Input：深度搜索：把“模型成本压降”相关的内部笔记和外部权威资料一起梳理