AIBuildAI Knowledge Base

The AIBuildAI Knowledge Base is a curated library of knowledge for building AI models of every kind, from computer vision and scientific AI to large language models. When an AI agent faces a design decision, such as which model architecture fits the task, which dataset to use, which training method to apply, or how to run a framework, it queries the knowledge base over MCP and reads what has worked, what has failed, and under which conditions. This repository is its source code: the retrieval service, the skill corpus, and the offline pipelines that build the corpus.

The service only retrieves. It returns material, never decisions, and the calling agent designs from what it reads, so building the index needs no LLM, only text splitting and embedding. A fixed snapshot of this corpus is served publicly at https://32.194.230.84/open, the default knowledge base address of AIBuildAI agents.

The corpus

The knowledge base holds one kind of knowledge, the skill: a SKILL.md with a description in its frontmatter and a markdown body, plus the reference files it carries, such as knowledge cards, paper summaries, modeling principles and code examples. A skill lives at data/corpus/plugins/<marketplace>/<plugin>/skills/<skill>/, and that triple is its address.

The corpus covers AI model building broadly, and post-training large language models is one area it treats in depth: a dedicated collection walks an agent through a post-training run, from choosing training data and a method to running it on a framework.

The 460 awesome-list skills span the major areas of AI:

Quick start

The public service at https://32.194.230.84/open serves this corpus, so an agent can use the knowledge base without hosting anything.

AIBuildAI agents. Mount it in the run config. The AIBuildAI LLM-Post-Train Agent connects to the public service by default:

mcps:
  kb: { type: reference }

Agents search it most when they are told to: say so in llm.system_instructions, as that repository's examples do.

Any MCP client. The service speaks MCP over Streamable HTTP at https://32.194.230.84/open/mcp. For example, in Claude Code:

claude mcp add --transport http kb https://32.194.230.84/open/mcp

The agent then gets four tools, named mcp__kb__* when the server is mounted as kb: list_plugin lists the plugins, search_skills searches the corpus, load_skill returns a skill with its files, and read_reference returns one file of a skill.

Host your own knowledge base (optional)

Build the index and serve it yourself to run on a private network or to serve a corpus you have extended.

git clone https://github.com/aibuildai-inc/aibuildai-knowledge-base.git && cd aibuildai-knowledge-base
python3.11 -m venv .venv && source .venv/bin/activate
pip install '.[server]'

export AIBUILDAI_KB_EMBED_MODEL=Qwen/Qwen3-Embedding-0.6B
RUNTIME_ROOT="$(PYTHONPATH=$PWD python -c 'from sourcing.utils.work_dir import runtime_root; print(runtime_root())')"

# build the search index, about 20 minutes for the full corpus on one GPU
python -m aibuildai_mcp.index --plugins-root data/corpus/plugins \
    --index-dir "$RUNTIME_ROOT/index" --model "$AIBUILDAI_KB_EMBED_MODEL"

# serve it at http://127.0.0.1:8000/mcp, with a health check at /healthz
AIBUILDAI_KB_DEVICE=cuda:0 bash deploy/serve_local.sh

The index build is incremental: it re-embeds only files whose content changed, and --force rebuilds everything. Restart the service after the corpus or the index changes. To point an AIBuildAI agent at your own service, set AIBUILDAI_KB_BASE_URL to its address; any other MCP client connects to its /mcp endpoint. The service has no authentication, so put it behind your own proxy if it must not be public.

Docker. deploy/deploy.sh builds the index, bakes it and the embedding model into an image, starts the container aibuildai-kb on port 8000, and checks that it answers a search. It needs Docker with the NVIDIA Container Toolkit.

CUDA_VISIBLE_DEVICES=0 AIBUILDAI_KB_EMBED_MODEL=Qwen/Qwen3-Embedding-0.6B bash deploy/deploy.sh

Growing the corpus

sourcing/ holds the offline pipelines that write corpus content. They call Claude through the Claude Agent SDK and need a different MCP version than the service, so install them in a separate environment with pip install '.[producers]'. See sourcing/README.md for each pipeline.

Project layout

aibuildai_mcp/        the retrieval service: index.py builds the index, service.py serves it
data/corpus/plugins/  the skill corpus
sourcing/             offline pipelines that write corpus content
deploy/               Dockerfile, deploy.sh, serve_local.sh
check.sh              pyright and ruff

Citation

If you use this knowledge base in your work, please cite:

@article{zhang2026aibuildai2,
    title={AIBuildAI-2: A Knowledge-Enhanced Agent for Automatically Building AI Models},
    author={Ruiyi Zhang and Peijia Qin and Qi Cao and Li Zhang and Pengtao Xie},
    year={2026},
    journal={arXiv},
    url={https://arxiv.org/abs/2605.27873}
}

License

Apache License 2.0

Community

Questions and discussion: join the AIBuildAI Community Slack or the AIBuildAI Community Discord. Use the issue tracker for bugs and feature requests.