Skip to main content
Idea generation turns the records you have collected in a notebook into structured research proposals. The Automated IdeaGen pipeline extracts knowledge points from those records, explores candidate ideas, applies two rounds of filtering, and outputs an organized Markdown document. For interactive, editor-based idea development, see Co-writer, which provides a freeform writing environment with AI editing and RAG-enhanced context.

Notebook-driven input

Works directly from your saved notebook records — no manual data entry needed.

Multi-stage filtering

Loose filter removes weak candidates; strict filter ensures quality before output.

Broad exploration

Generates at least 5 ideas per knowledge point from multiple research dimensions.

Structured output

Produces a Markdown document organized by knowledge point with ranked ideas.

​
How to use

1

Open the idea generation page

Navigate to http://localhost:3782/ideagen.
2

Select a notebook

Choose a notebook that contains records from previous sessions (solver outputs, research reports, question generation results, or co-writer documents).
3

Optionally add your thoughts

Enter any research directions, constraints, or focus areas you want the agent to prioritize.
4

Generate ideas

Click Generate Ideas. Progress updates stream in real time as each stage completes.
5

Review the output

The final Markdown document lists each knowledge point with its curated set of research ideas.

​
Automated IdeaGen pipeline

The pipeline has two components: a MaterialOrganizerAgent that extracts knowledge points from your notebook, and an IdeaGenerationWorkflow that turns those points into filtered research ideas.

​
MaterialOrganizerAgent

Reads the raw records from your selected notebook — which may include solve outputs, generated questions, research cache entries, and co-writer documents — and extracts a list of core knowledge points. Each knowledge point includes a name and a description derived from the record content. 
[
    {
        "knowledge_point": "Attention mechanism",
        "description": "Mechanism that assigns dynamic weights to input tokens..."
    },
    ...
]
Optional user_thoughts are passed into the extraction prompt to bias which knowledge points are surfaced.

​
IdeaGenerationWorkflow stages

Stage 1 — Loose filter Removes knowledge points that are obviously unsuitable for research idea generation (too narrow, too broad, already exhaustively studied, etc.). The remaining points proceed to exploration. Stage 2 — Explore ideas For each knowledge point that passes the loose filter, the agent generates at least 5 candidate research ideas. Ideas are produced from multiple innovation dimensions to maximize diversity. Stage 3 — Strict filter Each knowledge point’s idea list is evaluated strictly. The filter must keep at least 1 idea and eliminate at least 2. This ensures only the most promising, clearly differentiated ideas survive. Stage 4 — Generate Markdown The surviving ideas are assembled into a structured Markdown document grouped by knowledge point.

​
Output format

# Research Ideas

## Attention mechanism
Description: Mechanism that assigns dynamic weights to input tokens...

### Research Ideas
1. Sparse attention patterns for long-document summarization
2. Learned positional attention biases for code generation

## Knowledge point 2
...

​
Python API

You can run idea generation programmatically without the web interface. 
import asyncio
from src.agents.ideagen import IdeaGenerationWorkflow, MaterialOrganizerAgent
from src.core.core import get_llm_config

async def main():
    llm_config = get_llm_config()

    # Step 1: extract knowledge points from notebook records
    organizer = MaterialOrganizerAgent(
        api_key=llm_config["api_key"],
        base_url=llm_config["base_url"]
    )
    records = [...]  # list of notebook record dicts
    knowledge_points = await organizer.process(
        records,
        user_thoughts="Focus on efficiency improvements"
    )

    # Step 2: run the filtering and generation workflow
    workflow = IdeaGenerationWorkflow(
        api_key=llm_config["api_key"],
        base_url=llm_config["base_url"],
        progress_callback=print  # optional streaming updates
    )

    filtered_points = await workflow.loose_filter(knowledge_points)

    ideas_map = {}
    for point in filtered_points:
        ideas = await workflow.explore_ideas(point)
        filtered_ideas = await workflow.strict_filter(point, ideas)
        ideas_map[point["knowledge_point"]] = filtered_ideas

    markdown = await workflow.generate_markdown(filtered_points, ideas_map)
    print(markdown)

asyncio.run(main())

​
API endpoint

The module is also accessible via the REST API: 
POST /api/v1/ideagen/generate

{
  "notebook_id": "notebook_123",
  "user_thoughts": "Focus on efficiency improvements"
}
Response: 
{
  "success": true,
  "markdown": "# Research Ideas\n\n## ...",
  "knowledge_points": [...],
  "ideas_map": { "Attention mechanism": ["Idea 1", "Idea 2"] }
}

​
Relationship to co-writer

Automated IdeaGen produces a structured starting document. You can open that document in the Co-writer to expand individual ideas, rewrite sections with RAG-enhanced context, or generate a narrated audio version for review.
Idea generation requires at least one notebook with saved records. If your notebook is empty, complete some sessions in the solver, question generator, or deep research modules first and save results to a notebook.

Build docs developers (and LLMs) love

Get started for freeTalk to us