AF Deep Research

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. Built on top of agentfield runtime for orchestrating the agents.

One-Call DX

Trigger it with the af CLI (requires af ≥ 0.1.87) — it streams live progress and prints the result:

af call meta_deep_research.execute_deep_research --in '{"query": "What companies are investing in AI chips?"}'

Prefer raw HTTP? Hit the API directly with curl:

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 a structured research package — entities, relationships, evidence, and a full cited document:

{
  "entities": [
    {"name": "NVIDIA", "type": "Company", "summary": "Dominant AI chip maker, 80%+ datacenter GPU share"},
    {"name": "Sequoia Capital", "type": "Investor", "summary": "Early NVIDIA investor, major AI fund"},
    {"name": "H100", "type": "Technology", "summary": "Flagship AI training GPU, $30k+ ASP"}
  ],
  "relationships": [
    {"source": "AMD", "target": "NVIDIA", "type": "Competes_With"},
    {"source": "Sequoia Capital", "target": "NVIDIA", "type": "Invests_In"},
    {"source": "Microsoft", "target": "OpenAI", "type": "Partners_With"}
  ],
  "article_evidence": [{"facts": ["NVIDIA datacenter revenue reached $18.4B in Q4"], "quotes": ["We are seeing unprecedented demand - Jensen Huang"]}],
  "document": {"sections": ["...full hierarchical report with inline citations..."]},
  "metadata": {"iterations_completed": 3, "total_entities": 47, "total_relationships": 89, "final_quality_score": 0.91}
}

~10,000 agent invocations. Self-correcting research loops. One API call.

Quick start

git clone https://github.com/Agent-Field/af-deep-research.git && cd af-deep-research
cp .env.example .env          # Add OPENROUTER_API_KEY + at least one search provider key
docker compose up --build

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.

Run locally (without Docker)

You can run AF Deep Research directly as a Python app without the AgentField control plane managing async dispatch. It will execute synchronously.

1. Create a virtual environment and install dependencies:

python -m venv .venv && source .venv/bin/activate
pip install -e .

2. Set a writable working directory:

By default, AF Deep Research writes workspace files to /workspaces, which requires root access. Override this with the PR_AF_WORKDIR environment variable:

export PR_AF_WORKDIR=/tmp/pr-af-workdir

3. Configure your environment and start the agent:

cp .env.example .env   # Add OPENROUTER_API_KEY + at least one search provider key
python3 main.py

4. Submit a query (in a separate terminal):

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?"}}'

Output

The API returns a research package with five components:

Component	Contents
entities	Typed objects: Company, Investor, Founder, Technology, Market_Trend, Metric
relationships	Edges: Competes_With, Invests_In, Partners_With, Founded_By, Acquires
article_evidence	Facts and quotes with source article IDs
document	Hierarchical sections with inline citations and bibliography
metadata	iterations_completed, total_entities, total_relationships, final_quality_score

Entities

{
  "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

{
  "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"}
  ]
}

Evidence

{
  "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

{
  "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

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

How it works

What makes this different

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.

	Perplexity / ChatGPT	AF Deep Research
Research process	One-shot generation	Self-correcting loops that ask "what am I missing?"
Search coverage	Single query	Multiple parallel streams, each exploring a different angle
What you get back	Prose with inline links	Structured JSON: typed entities, mapped relationships, cited evidence
Context quality	Everything retrieved goes to the LLM	Two-tier filtering so only hyper-relevant content reaches the model
Relationship discovery	Mentioned if obvious	Multi-pass extraction: explicit → implied → indirect → emergent patterns
Scaling	Fixed behavior	Width and depth scale with query complexity
Integration	Copy-paste from chat	REST API, SSE streaming, webhook-ready
Deployment	SaaS only	Self-host with local LLMs, air-gapped option

From research to action: Contract-AF uses the same recursive multi-agent pattern to analyze legal contracts — agents spawn agents at runtime based on what they discover.

Why it works

• 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.

API

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}

Accessing response data

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()

Parameters

Parameter	What it does
research_focus	Depth (1-5). Higher goes deeper.
research_scope	Breadth (1-5). Higher casts wider net.
max_research_loops	How many iterative cycles to run.
tension_lens	balanced, bull, or bear perspective.
source_strictness	strict, mixed, or permissive source filtering.

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

Examples

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

Local LLMs

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.

Model options

Model	Cost
deepseek-chat-v3.1	$0.15/0.75 per 1M tokens
claude-sonnet-4	$3/$15 per 1M tokens
ollama/llama3.2	Free (local)
ollama/qwen2.5:72b	Free (local)

The stack

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.

Links

• Docs
• GitHub
• Discord

Contribute

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

---

Also built on AgentField

Contract-AF — Legal contract risk analyzer. Agents spawn agents at runtime. Adversarial review catches what solo LLMs miss.

Reactive Atlas — Turn any MongoDB collection into an AI intelligence layer. Config-driven, zero application code.

All repos →

_{Built by AgentField · Apache 2.0 · See what else we're building →}