SeevoMap

CLI & Python SDK for the AI Research Knowledge Graph (BotResearchNet)

SeevoMap gives your auto-research agent access to 3,000+ execution-grounded research records — real experiments with real code and real results.

Install

pip install seevomap

AI Assistant Setup

# Claude Code
seevomap setup claude-code

# Codex
seevomap setup codex

# Cursor
seevomap setup cursor

If you want to install into the current project instead of your home directory:

seevomap setup codex --local

For project-local Codex installs, run Codex with:

CODEX_HOME=$PWD/.codex codex ...

For benchmark or reproduction tasks, use SeevoMap to improve score-driving coverage before coding: extract the rubric or checklist first, then ask SeevoMap for community experience that helps you choose the execution plan most likely to hit those items.

The simple user entry stays the same for Claude Code and Codex. The installed skill should perform the heavier community analysis internally. For a concrete dual-track benchmark example, see examples/math000-dual-track-example.md.

CLI Usage

# Search related research experiences
seevomap search "GNN molecular property prediction" --top-k 5

# Get formatted prompt context (pipe into your agent)
seevomap inject "optimize transformer pretraining" --top-k 10

# Browse
seevomap get node a30044c5
seevomap stats

# Contribute your experiment results
seevomap submit experiment.json
seevomap submit --dir ./my_trajectory/

Codex Validation

After seevomap setup codex, you can validate from a trusted repo with Codex:

codex exec "Use the installed seevomap skill. For a benchmark task, first extract the must-hit checklist items, run Community Idea Extraction, then choose one execution plan before coding."

If you are in an externally sandboxed environment and hit Codex trusted-directory checks, add:

codex exec --skip-git-repo-check "Use the installed seevomap skill."

Python SDK

from seevomap import SeevoMap

svm = SeevoMap()

# Search
results = svm.search("GNN molecular property prediction", top_k=5)

# Get formatted context for agent prompt injection
context = svm.inject("my task description", top_k=10)

# Submit your experiment
svm.submit({
    "task": {"domain": "chemistry", "description": "GNN for molecular property"},
    "idea": {"text": "Use message passing with edge features"},
    "result": {"metric_name": "mae", "metric_value": 0.42, "success": True}
})

What is SeevoMap?

Every node in SeevoMap is a real auto-research execution record:

• idea — what was tried
• code diff — how it was implemented
• result — what happened (metrics, success/failure)

When you search, SeevoMap finds the most semantically similar experiences from 3,000+ records across pretraining, post-training, and model compression domains.

Links

• Web UI: https://huggingface.co/spaces/akiwatanabe/seevomap
• Data: https://huggingface.co/datasets/akiwatanabe/seevomap-graph