Searching Precisely#

Overview#

Web search pipeline that minimizes token consumption via local intent classification, semantic caching, credibility validation, and streaming fragment assembly.

架构分工：

• **宿主 AI（Host Agent）**负责实际的网页 GET / Search API 调用，返回原始 fragments
• 本 Skill 的脚本 负责前处理（intent 分类、query 改写、budget 控制、cache 查询）和后处理（credibility probe、stream 组装、cache 写入）

Core Rule: Always check the semantic cache first. Only invoke web search on a cache miss.

Pipeline Architecture#

text

Query → [Intent Parser] → [Query Rewriter] → [Budget Controller]
                                                     ↓
                                           [Semantic Cache] ──hit──→ Return
                                                     ↓ miss
                                           [Web Search]  (≤1500 tok)
                                                     ↓
                                           [Parallel Credibility Probe]
                                                     ↓
                                           [Stream Assembler] → [Write Cache]

Instructions#

When this skill activates, execute the pipeline below in order. Exit early at any step that produces a final answer — do not run later steps unnecessarily.

Note: Replace <placeholders> with actual runtime values. All arguments must be valid JSON strings.

---

Step 1 — Classify Intent#

Run via shell tool:

text

node scripts/intent-parser.js '<original_query>'

Extract intent and confidence from the JSON output.

If confidence < 0.5, default to intent = "web_search" and continue.

---

Step 2 — Initialize Budget#

text

node scripts/budget-controller.js init

Keep the returned state.remaining value. Abort any later step that would exceed it.

---

Step 3 — Check Semantic Cache#

text

node scripts/semantic-cache.js check '{"query":"<original_query>","intent":"<intent>"}'

• hit: true and similarity ≥ 0.85 → return result to the user. Pipeline complete. Skip all remaining steps.
• hit: false → continue to Step 4.

---

Step 4 — Rewrite Query#

text

node scripts/query-rewriter.js '{"intent":"<intent>","query":"<original_query>"}'

Use the returned subQueries array (max 3) for web search.

---

Step 5 — Web Search (host agent)#

Using your native search_web tool, search each sub-query from Step 4.

Collect result URLs and content fragments.

Always perform live search on a cache miss — never fabricate results.

---

Step 6 — Validate Source Credibility#

Extract up to 5 unique source URLs from Step 5. Run:

text

node scripts/parallel-probe.js '{"sources":[{"url":"<url1>"},{"url":"<url2>"}]}'

• verdict: "trust" → use directly
• verdict: "verify" → use with caution; flag in the answer
• available: false → discard that source

---

Step 7 — Score Credibility#

text

node scripts/credibility-arbiter.js '{"results":[<probe_results_array>]}'

If all sources score < 0.4, discard everything and tell the user no reliable source was found. Do not assemble.

---

Step 8 — Assemble Answer#

text

node scripts/stream-assembler.js '{"fragments":[<trusted_fragments>],"query":"<original_query>"}'

Return the answer field to the user.

If coherenceScore < 0.5, add a note that the result may be incomplete.

---

Step 9 — Write Cache#

text

node scripts/semantic-cache.js write '{"query":"<original_query>","intent":"<intent>","result":"<answer>"}'

This caches the result for future similar queries. Always run this step after a successful assembly.

Search Decision Rules#

SituationActionQuery answered by cache (similarity > 0.85)Return cache, skip web searchQuery about API docs / official specsWeb search with known domain hintQuery about current events / pricing / newsWeb search, no cache assumedQuery with ambiguous intentRewrite into 2-3 sub-queries, search each

Token Budget Defaults#

ComponentBudgetIntent parse50Cache check20Web search1500Credibility300Assembly500Total****2370

Common Mistakes#

• ❌ Calling web search before checking the semantic cache
• ❌ Skipping credibility check for unknown domains
• ❌ Assembling fragments from a single unverified source
• ❌ Fetching full page content when a HEAD probe + summary suffices

ClawHub Security Notes#

All permission declarations, dependencies, I/O scopes, and security properties are formally declared in the YAML frontmatter above ( permissions, dependencies, io, security_notes). Static analyzers should use those fields as the authoritative source.

Summary per script:

ScriptI/ONetworkShellintent-parser.jsnonenonenonequery-rewriter.jsnonenonenonestream-assembler.jsnonenonenonebudget-controller.jsnonenonenonecredibility-arbiter.jsnonenonenonesemantic-cache.js``~/.antigravity/search-cache/ R/Wnonenoneparallel-probe.jsreputation DB R (bundled)HEAD only, no uploadnone