Deep Research

Run a thorough, multi-agent investigation that produces a cited research brief.

Deep research is the flagship Bohr AI workflow. It now runs a planner-led, iterative architecture that combines evidence gathering, contradiction checks, reasoning validation, and citation verification before final delivery.

Usage

From the REPL:

/deepresearch What are the current approaches to mechanistic interpretability in LLMs?

From the CLI:

bohr deepresearch "What are the current approaches to mechanistic interpretability in LLMs?"

Both forms are equivalent. The workflow begins immediately and streams progress as agents discover and analyze sources.

How it works

The workflow runs through an orchestrated pipeline:

  1. Planning — research-planner creates the execution strategy, agent order, and stop criteria.
  2. Hypothesis + research fan-out — hypothesis frames testable lines of inquiry, then parallel researcher agents gather evidence across papers, docs, and code.
  3. Scoring + structure — evidence-scorer assesses claim strength and knowledge-graph captures reusable entity-relation structure.
  4. Adversarial checks — contradiction finds conflicts and edge cases; reasoning-validator checks whether conclusions logically follow from evidence.
  5. Draft + citation + review — writer synthesizes, verifier anchors citations, reviewer pressure-tests quality.
  6. Iteration loop — iteration-controller decides whether to continue another loop (Research -> Review -> Improve) until convergence criteria are met.

Default convergence policy targets high confidence and resolved critical contradictions before final output promotion.

Expected output

Deep research produces multiple durable artifacts:

  • outputs/.plans/<slug>.md — orchestration plan and verification ledger
  • outputs/<slug>-knowledge-graph.md — reusable structured knowledge
  • outputs/.drafts/<slug>-draft.md — synthesis draft
  • outputs/<slug>.md (or papers/<slug>.md) — final cited and verified deliverable
  • <slug>.provenance.md sidecar — source and verification accounting

The final brief typically includes:

  • executive summary
  • key findings by theme
  • contradictions and caveats
  • open questions
  • inline citations with source links

Customization

You can steer deep research by specifying:

  • scope — “focus on enterprise production deployments only”
  • time range — “use sources from 2023 onward”
  • evidence policy — “prioritize peer-reviewed + official docs”
  • convergence targets — “stop only when confidence > 0.85 and no critical contradictions”
  • output style — “brief for executives” vs “technical deep dive”

Real example

bohr deepresearch "Do long-context models reduce RAG complexity for regulated-domain QA in production?"

Typical output: a cited brief with trade-offs, contradiction handling, confidence-calibrated recommendations, and a reusable knowledge-graph artifact.