openclaw - 💡(How to fix) Fix [Ballast] Research agents — automated PubMed/web sourcing for SKU vetting [1 comments, 1 participants]

Overview

Build automated research agents that search the web and PubMed, extract study data, and produce professional vetting conclusions for supplement candidates. This automates the evidence research layer of the vetting pipeline.

Problem

Manual SKU vetting (issue #69867) requires reading PubMed papers and extracting:

• Mechanism of action
• Study design (RCT, observational, etc.)
• Effect size (Cohen's d, odds ratio, etc.)
• Sample size and population
• Dose used in study vs. product dose
• Limitations and conflicts of interest

This is time-consuming to do manually at scale. An agent pipeline can do the first-pass research, leaving the human to review conclusions rather than read raw papers.

Proposed agent design

Agent 1 — PubMed Scout

• Input: ingredient name + health category
• Searches PubMed via Entrez API (free, no key needed)
• Returns: top 5–10 most relevant studies by relevance + recency
• Extracts: title, authors, journal, year, abstract, PMID, DOI
• Output: structured study list with links

Agent 2 — Evidence Extractor

• Input: PubMed abstract (or full text if available via PMC)
• Uses LLM to extract:
  • Study type (RCT / meta-analysis / observational / case study)
  • Population (n, age, condition)
  • Intervention (dose, form, duration)
  • Primary outcome + effect size
  • Limitations acknowledged by authors
• Output: structured evidence record matching vetting.mjs schema

Agent 3 — Vetting Analyst

• Input: evidence records from Agent 2 + product spec from supplier
• Applies T1–T4 rubric + G1–G4 gates (from PRODUCT_VETTING_PROTOCOL.md)
• Produces verdict: approve / hitl_required / auto_reject
• Writes mechanism_one_sentence, nonclaims[], weak_evidence notes
• Flags: hallucinated URLs, under-dosed products, unsupported claims
• Output: vetting verdict matching runVetting() return schema

Agent 4 — Supplier Cross-check

• Input: approved ingredient name
• Searches CJ Dropshipping catalog for matching products
• Checks: spec sheet availability, US warehouse, dose per serving
• Returns: sourcing_url, spec_sheet_url, price, moq

Stack

• PubMed: NCBI Entrez E-utilities API (free)
• Web search: Tavily or Firecrawl (already in extensions/)
• LLM: chatWithFallback() from lib/ollama_cloud.mjs
• Output: feeds directly into vetting.mjs and initial_pass JSON

Integration point

These agents slot into the pipeline between Agent A (radar/trend discovery) and Agent B (store builder):

radar_candidates → Research Agents (PubMed + web) → vetting.mjs → B.5 product listing

Acceptance criteria

• Agent 1 returns ≥5 relevant PubMed studies for any of the 15 initial SKU candidates
• Agent 2 correctly extracts study type, n, dose, effect size from abstracts (spot-check 5 papers)
• Agent 3 produces verdicts matching manual vetting for ≥80% of initial SKU pass (#69867)
• All PubMed URLs verified live (HEAD fetch — no hallucinated PMIDs)
• Full pipeline run on one candidate takes < 60 seconds

Related

• #69867 Initial SKU research (manual baseline to validate against)
• #69866 vetting.mjs (consumes agent output)
• data/scripts/PRODUCT_VETTING_PROTOCOL.md
• extensions/firecrawl/, extensions/tavily/ (existing search tools)