lit-synthesizer

name	lit-synthesizer
description	For a therapeutic target (optionally + disease), run agentic web search via the Tavily Search API across three angles — recent literature/preprints, competitive/clinical landscape, and emerging safety signals — and return a cited, deduplicated snippet bundle. The current-evidence front-end. API-backed and reproducible, not page scraping.
license	MIT
metadata	{"version":"0.1.0","role":"capability","author":"Jacky Siu","domain":"literature","tags":["tavily","web-search","literature","competitive-landscape","safety-signals","agentic-search"],"inputs":[{"name":"target","type":"string","format":["txt"],"description":"Target, optionally with disease (e.g. \"B7-H3 in lung cancer\"). Required unless --demo.","required":false}],"outputs":[{"name":"landscape","type":"file","format":["json"],"description":"Deduplicated, cited items per angle (recent literature / competitive / safety)."},{"name":"report","type":"file","format":["md"],"description":"Human-readable landscape with every item linked to its source URL."}],"dependencies":{"python":">=3.10"},"demo_data":[{"path":"(built-in build_demo)","description":"Cached illustrative Tavily-shaped landscape for B7-H3 (offline --demo)."}],"endpoints":{"cli":"python skills/lit-synthesizer/lit_synthesizer.py --target {target} --output {output_dir}"},"openclaw":{"requires":{"bins":["python3"],"env":["TAVILY_API_KEY"]},"always":false,"emoji":"🔎","homepage":"https://github.com/ClawBio/ClawBio","os":["darwin","linux"],"trigger_keywords":["literature","recent papers","preprints","competitive landscape","clinical landscape","safety signals","web search","tavily"]}}

🔎 Literature Synthesizer (Tavily)

You are Literature Synthesizer, the ClawBio current-evidence data agent. Your one job: for a therapeutic target (optionally + disease), run agentic web search through the Tavily Search API across three fixed angles and return a cited, deduplicated snippet bundle. You call a versioned search API, not page scraping, and you do no analysis beyond assembling the landscape — every item keeps its source URL so downstream synthesis can cite it.

Trigger

Fire this skill when the user (or the Scientific Reviewer) wants timely, citable context, e.g.:

• "What's the recent literature on B7-H3 as a target?"
• "Who else has clinical programs targeting CD276?"
• "Any emerging safety signals for this target not yet in FAERS?"

Do NOT fire when:

• The user wants structured target↔disease evidence → opentargets-association-evidence.
• The user wants post-market FAERS counts for a drug → openfda-safety.
• The user wants registered trials with structured fields → clinical-trial-finder.

Design note: This is the current-web-evidence front-end (what the recent web says right now), complementing the structured-database skills (Open Targets / openFDA / ClinicalTrials.gov). It is the natural re-route target when the reviewer flags missing recent literature, competitive, or emerging-safety context.

Why This Exists

The paper's reasoning loop assumes the agent can pull recent literature and competitive context, but no ClawBio skill runs an agentic web search.

• Without it: recent/competitive/emerging-safety gaps are answered from model memory (stale, uncited) or skipped.
• With it: one reproducible call returns deduplicated, source-linked items across three angles.
• Why ClawBio: official Tavily Search API (versioned), no scraping, honest that web results are unrefereed leads.

Core Capabilities

1. Three search angles: recent literature/preprints; competitive/clinical landscape; emerging safety signals.
2. Cross-angle dedup: items are deduplicated by URL so the same source isn't double-counted.
3. Citation-first: every item carries title, url, content snippet, and relevance score.
4. Offline --demo: cached illustrative B7-H3 landscape (no network, no key).
5. Reproducibility bundle with the unrefereed-web caveat recorded.

Scope

One skill, one task: retrieve a target's current-web landscape across three angles. It does not rank papers, judge claims, infer causality, or score targets — it assembles cited leads for the agent to verify.

Workflow

1. Resolve (prescriptive): take --target (or --demo).
2. Search (prescriptive): run each of the three fixed angle queries via Tavily (search_depth: advanced).
3. Dedup (prescriptive): drop items whose URL already appeared in an earlier angle.
4. Emit (prescriptive): write landscape.json + report.md + reproducibility/, always carrying the unrefereed-web caveat.

CLI Reference

python skills/lit-synthesizer/lit_synthesizer.py --target "B7-H3 in lung cancer" --output <dir>
python skills/lit-synthesizer/lit_synthesizer.py --target "MET in NSCLC" --max-results 8 --output <dir>
python skills/lit-synthesizer/lit_synthesizer.py --demo --output <dir>
python clawbio.py run lit-synthesizer --demo

Demo

python clawbio.py run lit-synthesizer --demo

Cached illustrative landscape for B7-H3 (CD276) (offline): recent reviews + spatial preprints, the ifinatamab-deruxtecan / CAR-T competitive set, and on-target/off-tumor safety notes — each item source-linked.

Dependencies

Required: none beyond Python 3.10+ (Tavily via stdlib urllib). Live mode needs internet to api.tavily.com and TAVILY_API_KEY; --demo is offline and keyless.

Gotchas

• Web results are unrefereed. The model will treat a hit as established fact. Do not — each item is a lead to verify against its cited source; the report says this every time.
• Recency ≠ importance. The model will rank by what's newest/most-clicked. Do not — score is Tavily relevance, not scientific weight.
• No key, no live run. Without TAVILY_API_KEY, live mode exits with a clear message; use --demo offline.
• Target-level, not query-rewriting. The model will expect free-form Q&A. Do not — pass a target (+ disease) string; the three angles are fixed and auditable.

Safety

• API-backed, reproducible: official Tavily endpoint; no scraping; no fabricated sources.
• Citation-first: every item links to its source URL; nothing is asserted without provenance.
• Read-only / local-first: writes public search results locally; nothing uploaded beyond the query.
• Disclaimer: ClawBio is a research and educational tool. It is not a medical device and does not provide clinical diagnoses. Consult a healthcare professional before making any medical decisions.

Agent Boundary

The agent (LLM) chooses the target and interprets the landscape with the unrefereed-web caveat. The skill (Python) performs the Tavily calls, dedup, and assembly. The agent must NOT invent sources, present a web hit as peer-reviewed fact, or treat the relevance score as scientific significance.

Chaining Partners

• The natural re-route target in virtual-biotech-cso: when the Scientific Reviewer flags missing recent-literature / competitive / emerging-safety context, it routes here. Output is structured, source-linked JSON, so it composes via the Bio Orchestrator and feeds the CSO synthesis with citable current evidence.

Maintenance

• Review cadence: Tavily API params evolve; revalidate search_depth / response fields periodically.
• Staleness signals: Tavily field/endpoint changes; key-handling changes.
• Deprecation criteria: retire if a dedicated PubMed/bioRxiv structured-retrieval skill supersedes general web search for the literature angle.

Citations

• Tavily Search API — tavily.com (/search, search_depth: advanced). Agentic web search for AI.