Deep research mode — multi-agent, evidence-backed
The AMRS (Adaptive Multi-Agent Research Swarm) pipeline runs 4 specialized agents in parallel — each searching a different angle, cross-validating findings, and assembling a structured evidence graph. Research that takes hours of manual work completes in minutes.
4 Agent Types Running in Parallel
Agents communicate via tokio channels. Each runs independently; the synthesis agent assembles the final report once all others complete.
Search Agent
Dispatches the primary query and multiple semantic variants across all backends. Collects the initial evidence pool.
Verification Agent
Takes each claim from the search agent and independently searches for corroborating or contradicting evidence. Flags contradictions.
Deep Dive Agent
For the highest-scoring sources, follows links, extracts referenced papers, and builds a second-degree evidence graph.
Synthesis Agent
Combines findings, resolves contradictions using the RAR protocol, ranks evidence by consensus confidence, and structures the final report.
RAR: Retry-and-Refine Self-Correction
The synthesis agent runs 5 reflection checkpoints before returning the final report. Low-confidence sections are rewritten; contradictions are resolved or flagged.
Start a Research Task
curl -X POST https://api.fetchium.com/v1/research \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"query": "State of AI inference optimization in 2025",
"depth": "standard",
"max_sources": 20,
"citation_format": "apa",
"token_budget": 16384
}'