qa10devteam/behive/hermes

BeHive Deep Research

Overview

BeHive is a deep research engine that extracts structured, scored claims from any topic. Unlike simple web search, it:

1. Decomposes topics into research axes
2. Fetches 1000+ sources via 8-layer stealth drones (70+ APIs)
3. Extracts typed claims with per-claim quality scores (0.0–1.0)
4. Builds entity relationship graphs (Neo4j)
5. Produces synthesized reports with inline citations

This skill connects Hermes Agent to a running BeHive instance.

When to Use

• User says "research X deeply", "gather intelligence on Y", "find everything about Z"
• User needs verified facts with sources, not LLM-generated summaries
• User wants claims they can cite, filter by quality, or query later
• User asks to build a knowledge base on a topic over time
• User says "run a mission on X", "scale 100", "deep dive"

Don't Use For

• Simple factual lookups (use web_search)
• Single-page content extraction (use web_extract)
• Real-time news (BeHive takes 5-15 min per mission)

Prerequisites

BeHive API running at http://localhost:8091 or configured endpoint.

Verify:

curl -s http://localhost:8091/health | jq .

If using MCP (recommended), configure in ~/.hermes/config.yaml:

mcp_servers:
  behive:
    url: http://localhost:8090/mcp
    transport: streamable-http

Quick Research (MCP)

If BeHive MCP is configured, use MCP tools directly:

mcp_behive_research_topic(request={"query": "NVIDIA GPU market 2025-2026", "depth": 3, "force": true})

Then poll:

mcp_behive_mission_status(job_id="<returned_id>")

Get report:

mcp_behive_get_report(job_id="<id>", format="markdown")

Search past knowledge:

mcp_behive_search_knowledge(query="NVIDIA revenue", limit=20)

Research via REST API (terminal)

Start a Mission

curl -s -X POST http://localhost:8091/research \
  -H "Content-Type: application/json" \
  -d '{"topic": "EU AI Act enforcement mechanisms 2026", "scale": 30, "depth": 3}' | jq .

Scale guide:

• 15 — quick scout (2-3 min, ~50 claims)
• 30 — standard (5-10 min, ~200-500 claims)
• 100 — deep (15-30 min, ~800+ claims)
• 300 — exhaustive (45-90 min, ~2000+ claims)

Poll Status

curl -s http://localhost:8091/research/<mission_id> | jq .status,.phase,.progress

Phases: scout → harvest → process → synth → done

Get Report

curl -s http://localhost:8091/research/<mission_id>/report | jq .

Search Claims

curl -s "http://localhost:8091/search?q=NVIDIA+revenue&limit=20" | jq .

List Missions

curl -s http://localhost:8091/missions | jq '.[] | {id, topic, status, avg_quality, total_claims}'

SSE Streaming (real-time progress)

curl -N http://localhost:8091/research/<mission_id>/events

Returns Server-Sent Events with phase transitions, progress %, and claim counts.

Quality Interpretation

Knowledge Graph Queries

# Get entities
curl -s "http://localhost:8091/graph/entities?limit=50" | jq .
 
# Get relationships for an entity
curl -s "http://localhost:8091/graph/entity/NVIDIA/relationships" | jq .
 
# Network stats
curl -s http://localhost:8091/intelligence/stats | jq .

Workflow: Research → Report → Deliver

1. Start mission with appropriate scale
2. Monitor via SSE or polling (every 30s)
3. On completion, fetch report
4. Format key findings for user
5. Offer to search specific claims or explore the knowledge graph

Common Pitfalls

1. Using web_extract on localhost — Hermes blocks private IPs via web_extract. Always use terminal() + curl for BeHive API calls.
2. Scale too high for simple topics — Scale 30 is sufficient for most queries. Scale 300 takes 45+ minutes and may hit rate limits.
3. Not waiting for completion — Missions are async. Always poll or stream until status=done.
4. Confusing mission_id formats — IDs look like hive_1785227949_815112. Copy exactly from the start response.
5. Expecting real-time results — Even scale 15 takes 2-3 minutes. Set user expectations.

Benchmarks (real, honest)

Hardware: EC2 g6.24xlarge, 4× NVIDIA L4, Bedrock Claude Haiku + Sonnet.

Verification Checklist

• [ ] curl http://localhost:8091/health returns 200
• [ ] MCP configured in config.yaml (if using MCP path)
• [ ] Mission started and job_id captured
• [ ] Status polled until done or error
• [ ] Report delivered to user in readable format