Autonomous research backend for AI applications.

Early Preview · APIs may change. Feedback welcome.

A research API that questions itself. Submit a query. The system spawns parallel agents, evaluates what it found against quality thresholds, generates new sub-queries when gaps are detected, and runs another cycle. The architecture orchestrates ~10,000 logical agent invocations per research. This is an AI backend, not a chat interface.

git clone https://github.com/Agent-Field/af-deep-research.git && cd af-deep-research
cp .env.example .env
# Add API keys to .env
docker-compose -f docker-compose.hub.yml up -d

curl -X POST http://localhost:8080/api/v1/execute/async/meta_deep_research.execute_deep_research \
  -H "Content-Type: application/json" \
  -d '{"input": {"query": "What companies are investing in AI chips?"}}'

Returns execution_id. Stream progress via SSE, fetch results when complete.

Open localhost:8080/ui to watch the workflow live.

The API returns a research package with five components:

{
  "entities": [
    {"name": "NVIDIA", "type": "Company", "summary": "Dominant AI chip maker, 80%+ datacenter GPU share"},
    {"name": "Jensen Huang", "type": "Founder", "summary": "CEO of NVIDIA since 1993"},
    {"name": "Sequoia Capital", "type": "Investor", "summary": "Early NVIDIA investor, major AI fund"},
    {"name": "H100", "type": "Technology", "summary": "Flagship AI training GPU, $30k+ ASP"},
    {"name": "AI Chip Market", "type": "Market_Trend", "summary": "Projected $300B by 2030"}
  ]
}

{
  "relationships": [
    {"source": "AMD", "target": "NVIDIA", "type": "Competes_With", "description": "Direct competition in AI accelerators"},
    {"source": "Sequoia Capital", "target": "NVIDIA", "type": "Invests_In", "description": "Series A investor, 1993"},
    {"source": "NVIDIA", "target": "Mellanox", "type": "Acquires", "description": "$7B acquisition for networking"},
    {"source": "Microsoft", "target": "OpenAI", "type": "Partners_With", "description": "$10B strategic investment"}
  ]
}

{
  "article_evidence": [{
    "article_id": 1,
    "facts": [
      "NVIDIA datacenter revenue reached $18.4B in Q4",
      "H100 backlog extends into late 2025"
    ],
    "quotes": ["We are seeing unprecedented demand - Jensen Huang"]
  }]
}

{
  "document_title": "AI Chip Investment Landscape",
  "executive_summary": "The AI chip market is experiencing...",
  "sections": [
    {"title": "Market Dynamics", "content": "The accelerator market reached $45B [1]..."},
    {"title": "Competitive Landscape", "content": "NVIDIA maintains 80%+ share [2]..."}
  ],
  "source_notes": [
    {"citation_id": 1, "title": "AI Chip Report 2024", "domain": "reuters.com"}
  ]
}

{
  "metadata": {
    "iterations_completed": 3,
    "total_entities": 47,
    "total_relationships": 156,
    "total_sources": 89,
    "final_quality_score": 0.82
  }
}

Traditional research: one query → retrieve → summarize → done.

This system: fan out → filter down → synthesize → find gaps → fan out again.

Multiple parallel streams explore different angles simultaneously. Each stream filters aggressively—only hyper-relevant evidence survives. Cross-stream synthesis finds patterns. Gap analysis identifies what's missing. New targeted streams spawn. The cycle repeats until quality threshold is met.

• Context pollution kills quality. When you dump 50 web pages into an LLM, important facts get buried and the model hallucinates connections. We filter aggressively—hash dedup, semantic dedup, relevance scoring—so only the best evidence reaches synthesis.

• Width AND depth. Traditional tools go broad OR deep. This system scales both: research_scope controls parallel streams, research_focus controls analysis depth. Simple queries get fewer streams; complex queries spawn more.

• Quality-driven, not count-driven. The system doesn't run exactly 3 iterations. After each pass it asks: what entities are missing? what relationships are unclear? what claims lack evidence? If gaps exist and are addressable, it continues. If quality threshold is met, it stops.

• Cross-stream patterns. The synthesis layer doesn't summarize streams separately—it finds connections between them. A market trend that explains a technical pivot. A team background that validates a GTM strategy.

Submit: POST /api/v1/execute/async/meta_deep_research.execute_deep_research

Stream progress: GET /api/ui/v1/workflows/{execution_id}/notes/events (SSE)

Fetch results: GET /api/v1/executions/{execution_id}

response = fetch_results(execution_id)

# Graph data
entities = response["research_package"]["entities"]
relationships = response["research_package"]["relationships"]

# Evidence for audits
evidence = response["research_package"]["article_evidence"]

# Document for rendering
sections = response["research_package"]["document"]["sections"]
bibliography = response["research_package"]["document"]["source_notes"]

# Quality gates
if response["metadata"]["final_quality_score"] < 0.7:
    trigger_human_review()

Set tension_lens: "bear" for risk-focused analysis. Set source_strictness: "strict" to filter to reputable sources only.

Investment research pipeline

# Risk analysis for due diligence
curl -X POST http://localhost:8080/api/v1/execute/async/meta_deep_research.execute_deep_research \
  -H "Content-Type: application/json" \
  -d '{"input": {"query": "Rivian competitive position and financial risks", "tension_lens": "bear"}}'

# Response feeds into: risk scoring model, portfolio dashboard, analyst report generator

Knowledge graph builder

# Extract entities and relationships for graph database
curl -X POST http://localhost:8080/api/v1/execute/async/meta_deep_research.execute_deep_research \
  -H "Content-Type: application/json" \
  -d '{"input": {"query": "AI chip supply chain: manufacturers, suppliers, customers"}}'

# entities[] → Neo4j nodes
# relationships[] → Neo4j edges
# Now queryable: "Show me all companies 2 hops from NVIDIA"

Competitive intelligence system

# Track competitor moves across multiple dimensions
curl -X POST http://localhost:8080/api/v1/execute/async/meta_deep_research.execute_deep_research \
  -H "Content-Type: application/json" \
  -d '{"input": {"query": "How is AMD positioning against NVIDIA in datacenter AI?", "research_scope": 5}}'

# article_evidence[] → feed alerting system
# document → auto-generate weekly competitor brief

Compliance research

# Audit-ready research with source tracking
curl -X POST http://localhost:8080/api/v1/execute/async/meta_deep_research.execute_deep_research \
  -H "Content-Type: application/json" \
  -d '{"input": {"query": "ESG practices of major lithium mining companies", "source_strictness": "strict"}}'

# Each fact links to source_notes[] with URL, title, domain
# Audit trail shows exactly where each claim came from

To keep everything on your network:

OLLAMA_BASE_URL=http://host.docker.internal:11434
DEFAULT_MODEL=ollama/llama3.2

No telemetry. Your data stays on your infrastructure.

Runs on AgentField, open-source infrastructure for production AI agents. Workflows run for 16+ minutes without timeout. Progress streams via SSE. Results persist. Audit trails are cryptographically signed for compliance.

• Docs
• GitHub
• Discord

This is an early preview. We're actively developing and want feedback. File issues, open PRs, or come chat in Discord.

_{Built by AgentField · Apache 2.0}