mofa-org/mofa-skills/mofa-research

mofa-research

A DOT-based deep research pipeline that dynamically plans N search angles via LLM, executes them in parallel, then analyzes and synthesizes sequentially:

plan_and_search ──(N dynamic workers)──→ analyze ──→ synthesize
                                          (strong)    (strong, goal gate)

1. Plan — LLM generates 4-6 search angles tailored to the query (strong model)
2. Search — Spawns N search agents concurrently (cheap model, web tools)
3. Analyze — Cross-references all search outputs, organizes by subtopic (strong model)
4. Synthesize — Produces a structured report with citations (strong model, goal gate)

Usage

run_pipeline(pipeline="mofa-research/deep_research", input="<research topic>")

The pipeline file is at ~/.octos/skills/mofa-research/deep_research.dot.

How it works

The plan_and_search node uses handler="dynamic_parallel":

1. The planner LLM call generates a JSON array of {task, label} objects
2. Each task becomes a synthetic Codergen worker node with the worker_prompt template
3. All workers run concurrently via futures::join_all
4. When all complete, their outputs are merged and fed to the analyze node
5. Fallback: if the planner fails or returns <2 tasks, 3 generic angles are used

Compared to the old static parallel approach (4 hardcoded angles), dynamic_parallel:

• Adapts the number and type of search angles to each query
• Generates cross-language angles automatically
• Can produce 4-8 angles depending on topic complexity

Performance

With dynamic parallel search agents (~300s each):

• Sequential (old): ~1200s + analyze + synthesize = ~1800s
• Dynamic parallel (new): ~300s + analyze + synthesize = ~900s (2x overall speedup)

Data flow between nodes

• plan_and_search → analyze: All N worker outputs are merged with ## Label headers and --- separators
• analyze → synthesize: The analysis is passed as-is to the synthesis node

Each worker runs an independent agent with fresh context plus the original user input.

Customization

Copy deep_research.dot and edit:

• prompt — change the planning prompt to control how angles are generated
• worker_prompt — template for each worker (use {task} placeholder)
• planner_model / model — change which ProviderRouter keys to use
• tools="deep_search,read_file" — restrict which tools each worker can access
• max_tasks="8" — cap on number of dynamic tasks
• timeout_secs="600" — per-worker timeout
• Add intermediate nodes (e.g. fact_check between analyze and synthesize)