web_search_xray

Web Search Xray

web_search_xray is built from the same core workflow as maratha_web_search: generate diverse queries, search in parallel, deduplicate results, deep dive into the strongest sources, and produce a concise report with citations.

The difference is that web_search_xray adds a stricter evidence pass: it explicitly ranks sources by authority, recency, and directness so the final answer is easier to audit.

Use This Skill When

• The user asks for a web search, verification, or latest information.
• The user wants direct links or source-backed claims.
• The topic benefits from broad query coverage before narrowing to the strongest results.

Step 1: Generate Queries

Analyze the user's topic and craft 5-10 diverse search queries approaching it from different angles:

• Direct: Use the most straightforward phrasing of the topic
• Synonyms: Rephrase with alternative keywords, synonyms, and related terminology
• Specificity ladder: Create both broader and narrower versions of the query
• Temporal: Add year qualifiers (for example, 2026, latest, recent) where relevant
• Source-targeted: Append terms like site:github.com, research paper, official docs, tutorial, comparison, vs, benchmark as appropriate
• Question forms: Rephrase as how to, what is, why does, best practices for
• Technical depth: Add terms like internals, architecture, implementation, deep dive

Start with at least 5 queries. For ambiguous or fast-moving topics, use 8-10.

Step 1.5: Advanced Operator Mode (Optional)

If the broad-query pass is noisy, ambiguous, or too shallow, switch into an operator-driven search pass.

Use this mode when you need to:

• find papers on arxiv.org
• find blog posts from a specific site or class of sites
• find PDFs, docs, benchmarks, changelogs, or tutorials
• exclude low-value results and tighten precision quickly

Operator Toolbox

Use the following operators when the search engine supports them:

• "exact phrase": exact-match a phrase
• OR or |: allow alternatives
• AND: require both terms
• -term: exclude a term or phrase
• *: wildcard within a phrase
• ( ... ): group alternatives
• site:: restrict to a domain or site
• filetype: or ext:: restrict to file types such as pdf
• intitle:: require a term in the title
• allintitle:: require multiple title terms
• inurl:: require a term in the URL
• allinurl:: require multiple URL terms
• intext:: require a term in the body text
• allintext:: require multiple body terms
• before:: restrict to results before a date
• after:: restrict to results after a date
• related:: find related domains
• define:: definition lookup

Use with caution because support can be inconsistent:

• AROUND(X): proximity search
• #..#: numeric range search
• source:: news-source targeting

Avoid relying on deprecated or weak operators unless you have a concrete reason and verified they still work in your target engine.

Query Recipes

Find Papers On arXiv

Use site:arxiv.org first, then tighten with title, URL, phrase, and time filters.

Examples:

• "reward hacking" site:arxiv.org
• "mixture of experts" site:arxiv.org after:2024-01-01
• site:arxiv.org (abs OR pdf) "instruction following"
• site:arxiv.org inurl:abs "reinforcement learning with verifiable rewards"
• site:arxiv.org intitle:"reward hacking"
• site:arxiv.org filetype:pdf "transformer scaling"

For known paper IDs, search exact identifiers:

• site:arxiv.org "2508.04632"
• site:arxiv.org "2602.17004"

Find Blogs In General

Use site:, intitle:, inurl:, phrase matching, and exclusions together.

Examples:

• "reward hacking" (blog OR post OR essay)
• intitle:"reward hacking" (blog OR essay)
• inurl:blog "mixture of experts"
• (site:substack.com OR site:medium.com) "instruction following"
• (site:blog.google OR site:openai.com OR site:anthropic.com) "reinforcement learning"
• "search operators" -site:ahrefs.com

Find Technical Docs, PDFs, And Tutorials

Examples:

• "X API" (docs OR reference OR quickstart)
• site:docs.x.com "follows" after:2024-01-01
• "reward model" filetype:pdf
• "Mixture of Experts" (tutorial OR guide OR benchmark)
• intitle:benchmark "long context" LLM
• (changelog OR release notes) "vLLM"

Find Broad Web Coverage While Keeping Precision

Generate a query pack with one broad, one constrained, and one exclusion variant:

• Broad: "topic"
• Constrained: "topic" site:example.com OR site:example.org
• Exclusion: "topic" -reddit -pinterest -linkedin

Operator Strategy Rules

• Never use only operator-heavy queries. Pair them with plain-language broad queries.
• Use site: early when the user already has a likely source domain.
• Use after: and before: for unstable topics, releases, APIs, laws, benchmarks, and news.
• Use filetype:pdf for papers, reports, whitepapers, and slide decks.
• Use intitle: and inurl: when topical matching in the snippet is too weak.
• Use -term aggressively to remove SEO spam, mirrors, aggregators, and irrelevant brands.
• When searching for papers, include both title-based and ID-based variants.
• When searching for blogs, include both domain-specific and cross-domain variants.

Step 2: Search In Parallel

Send all queries as parallel web-search tool calls in a single message. Do not run them sequentially.

Step 3: Aggregate And Deduplicate

Once all searches return:

1. Remove duplicate URLs that appear across multiple searches.
2. Prioritize results that appear in multiple searches because they have a stronger relevance signal.
3. Group results by type, such as official docs, blog posts, discussions, papers, or tutorials.

Step 3.5: Rank Sources

Before opening pages, label candidates as:

• Primary: official documentation, original papers, maintainer posts, vendor docs
• Secondary: credible technical blogs, benchmark roundups, news summaries
• Tertiary: forum chatter, reposts, low-context summaries

Open Primary sources first unless they are clearly stale or off-topic.

Step 4: Deep Dive Into Top Results

Use page-fetching or page-open tools on the most promising results. Prioritize:

• Primary sources and official documentation
• Results whose snippets are insufficient but look highly relevant
• Pages likely to contain detailed technical information

During deep dives:

• Capture exact dates for unstable facts
• Note whether a claim is directly stated or inferred
• Prefer product docs, API references, papers, or changelogs over commentary

Step 4.5: Validate The Answer Surface

Before writing the response, check:

1. Does the answer rely on at least one primary source when one exists?
2. Are time-sensitive claims tied to a concrete date?
3. Are conflicts between sources explained instead of silently averaged away?
4. Is each recommendation or conclusion backed by a source that actually supports it?

Step 5: Write The Report

Structure findings as follows:

1. TL;DR: 2-3 sentence executive summary
2. Detailed sections: Organize findings under logical headers with specific facts and source-backed claims
3. Conflicts: Note where sources disagree
4. Sources: List all relevant URLs as markdown links at the end

When appropriate, add a short Evidence Quality line explaining whether the answer is based on official docs, papers, benchmarks, or community reports.

Rules

• Always search broadly first, then narrow to the best sources.
• Always prefer parallel search and fetch calls over sequential ones.
• Every claim must be traceable to a source URL.
• Keep source attribution explicit when the topic is fast-moving or high-stakes.
• Make temporal uncertainty visible. If a fact may have changed, verify it instead of guessing.
• Do not pad the result with weak sources when a smaller set of strong sources is sufficient.