Tavily

Tavily (Web Search / Extract / Research)

Prereqs

• Ensure TAVILY_API_KEY is set in the Hermes environment (commonly ~/.hermes/.env).
• Do not hardcode or paste API keys into chat logs. See references/bp-api-key-management.md.

Security Notes

• The bundled CLI (scripts/tavily.py) reads only TAVILY_API_KEY from the environment and only sends requests to https://api.tavily.com.
• Prefer search then extract over include_raw_content on search to keep outputs small and reduce accidental data exposure.

Quick Reference

Use the terminal tool to run the bundled CLI script (prints JSON). SKILL_DIR is the directory containing this SKILL.md file.

# Search (general)
python3 SKILL_DIR/scripts/tavily.py search --query "latest OpenAI API changes" --max-results 5

# Search (news) with recency filter
python3 SKILL_DIR/scripts/tavily.py search --query "latest OpenAI API changes" --topic news --time-range week --max-results 5

# High-precision search (more cost/latency)
python3 SKILL_DIR/scripts/tavily.py search --query "OpenAI API rate limits March 2026" --search-depth advanced --chunks-per-source 3 --max-results 5

# Search + answer (still cite URLs from results)
python3 SKILL_DIR/scripts/tavily.py search --query "What is X?" --include-answer basic --max-results 5

# Extract (targeted chunks; prefer this over include_raw_content on search)
python3 SKILL_DIR/scripts/tavily.py extract --url "https://example.com" --query "pricing" --chunks-per-source 3 --format markdown

# Research (creates task + polls until complete)
python3 SKILL_DIR/scripts/tavily.py research --input "Summarize the EU AI Act enforcement timeline. Provide numbered citations." --model auto --citation-format numbered --max-wait-seconds 180

Use returned results[].url fields as citations/sources in your final answer.

No-Script Option (curl)

Use Tavily directly via curl (same endpoints, no bundled script):

curl -s "https://api.tavily.com/search" \
  -H "Authorization: Bearer <TAVILY_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{"query":"latest OpenAI API changes","topic":"news","time_range":"week","max_results":5}'

curl -s "https://api.tavily.com/extract" \
  -H "Authorization: Bearer <TAVILY_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{"urls":"https://example.com","query":"pricing","chunks_per_source":3,"extract_depth":"basic","format":"markdown"}'

Procedure

1. Turn the user request into a focused search query (keep it short, ideally under ~400 chars). Split multi-part questions into 2-4 sub-queries.
2. Choose topic:
   • general for most searches
   • news for current events (prefer also setting time_range or date range)
   • finance for market/finance content
3. Choose search_depth:
   • Start with basic (1 credit) unless you need higher precision.
   • Use advanced (2 credits) for high-precision queries; use chunks_per_source to control snippet volume.
4. Keep max_results small (default 5) and filter by score + domain trust.
5. For primary text, run extract on 1-3 top URLs:
   • Provide an extract --query ... --chunks-per-source N to avoid dumping full pages into context.
6. For synthesis across multiple subtopics with citations, run research and poll until status=completed.

Pitfalls

• include_raw_content on search can explode output size; prefer the two-step flow: search then extract.
• auto_parameters can silently pick search_depth=advanced (2 credits). Set --search-depth explicitly when you care about cost.
• exact_match is restrictive; wrap the phrase in quotes inside --query and expect fewer results.
• country boosting is only available for topic=general.
• On failures, keep the request_id from responses for support/debugging.

Verification

• Check credits/limits: python3 SKILL_DIR/scripts/tavily.py usage
• Add --include-usage on search/extract if you want per-request usage info.

References

• references/search.md
• references/extract.md
• references/research.md
• references/research-get.md
• references/bp-search.md
• references/bp-extract.md
• references/bp-research.md
• references/bp-api-key-management.md
• references/usage.md