Openclaw Glasses

OpenClaw Glasses

See wider. Rank smarter. Answer with context.

OpenClaw Glasses is a search layer for OpenClaw. It starts with ordinary web lookup, then adds multi-source retrieval, intent-aware reranking, adaptive weighting, optional thread-pulling research, Chinese-query optimization, and finance-aware realtime prioritization.

OpenClaw Glasses 是一个给 OpenClaw 用的“搜索层 / 增强检索层”。它不是简单叠加几个搜索源，而是把多源召回、意图感知排序、权重自适应、链式追踪、中文优化、金融实时优先级整合成一条完整检索链，让结果更接近“先找对，再排对，最后答对”。

Public-facing summary

OpenClaw Glasses extends OpenClaw's native web tools into a smarter retrieval stack:

• multi-source search for broader recall and lower single-source bias
• intent-aware search for factual lookups, status/news, comparisons, tutorials, and exploratory research
• adaptive weighting so ranking changes with query type instead of using one fixed recipe
• thread-pulling / follow-up research for issues, discussions, and linked references
• Chinese-query optimization with CJK-aware matching and source weighting
• finance-aware realtime prioritization for stocks, indices, forex, and crypto quotes

OpenClaw Glasses 会把 OpenClaw 原生 web tools 扩展成一条更完整的检索链：

• 多源搜索：扩大召回面，减少单一来源偏差
• 意图感知检索：区分事实查询、状态更新、新闻、对比、教程、探索式研究
• 权重自适应：不同问题走不同排序逻辑，而不是一套固定权重打天下
• 链式追踪 / 深挖：遇到 issue、讨论帖、引用链时可以继续往下追
• 中文搜索优化：针对中文查询做 CJK-aware 匹配与中文友好源加权
• 金融实时增强：对股票、指数、外汇、加密资产等实时价格问题给出更稳的优先级

Example triggers

• "帮我查一下 OpenClaw 最新进展，并按可靠性排序"
• "Compare Bun vs Deno for production backend use"
• "AAPL 最新股价"
• "BTC 实时价格和 24h 涨跌"

Quick start

1. Use OpenClaw's built-in web_search as the agent-facing source when available.
2. Use scripts/search.py to aggregate additional providers and rerank results.
3. For status / exploratory / comparison work, prefer multi-query retrieval and intent scoring.
4. For finance price queries, let the finance-aware path boost Alpha Vantage and Binance results.

What this skill adds

• Intent-aware search modes: factual, status, comparison, tutorial, exploratory, news, resource
• Multi-source aggregation: Exa, Tavily, Grok, Gemini, Kimi
• Chinese-query optimization:
  • CJK-aware keyword matching instead of space-splitting only
  • modest boosts for Chinese-friendly sources when the query is in Chinese
• Finance-aware weighting:
  • boosts Alpha Vantage for stocks / ETFs / forex / index proxies
  • boosts Binance for crypto realtime quotes
• Optional GitHub thread-pulling and reference extraction for deeper research

Workflow

1. Pick the mode by intent

• Factual / tutorial → answer or light deep
• Status / news / comparison / exploratory → deep
• Resource finding → fast
• Finance realtime queries → fast for direct quote lookups, deep when combining quote + broader context

For intent examples and phrasing cues, read references/intent-guide.md.

2. Run the aggregator

Basic:

python3 scripts/search.py "query" --mode deep --intent exploratory --num 5

Multi-query comparison:

python3 scripts/search.py \
  --queries "Bun vs Deno" "Bun advantages" "Deno advantages" \
  --mode deep \
  --intent comparison

Finance quote:

python3 scripts/search.py "BTC 实时价格" --mode deep --intent status --source alpha-vantage,binance,gemini,kimi,tavily

3. Synthesize by topic, not by provider

• Answer first, then cite
• Group by themes or findings
• Call out conflicts explicitly
• Treat single-source or older claims more cautiously

Scripts

scripts/search.py

Primary multi-source retrieval and reranking entrypoint.

Capabilities:

• intent-aware scoring
• multi-query execution
• provider fusion
• Chinese-query weighting
• finance-aware realtime boosts
• optional extract-refs integration

scripts/fetch_thread.py

Deep-fetch GitHub issues / PRs or generic pages to extract structured references.

scripts/chain_tracker.py

Recursive thread-pulling / follow-up exploration with relevance gating.

scripts/relevance_gate.py

Batch relevance filtering for candidate links.

References

• references/intent-guide.md — intent cues and search-mode guidance
• references/authority-domains.json — authority weighting rules
• references/research-light-regression-samples.md — research-light behavior examples

Configuration notes

Do not hardcode secrets in the skill.

Expected runtime configuration:

• search provider keys via environment or a local credentials file
• optional reuse of OpenClaw's existing web-search provider config
• finance sources should remain optional; degrade gracefully if unavailable

Publishing / safety

Before packaging or publishing:

• remove all plaintext secrets
• remove machine-specific notes, personal paths, and private identifiers
• verify that examples and docs contain no local credentials or private data
• run the validator / packager before publishing