Code Docs Builder

Code Docs Builder#

Overview#

Code Docs Builder UI

This Solution Blueprint generates technical documentation directly from a GitHub repository using AMD Inference Microservices (AIMs). This addresses the common problem of missing or outdated documentation by keeping generated documentation aligned with the actual implementation.

The system analyzes a Git repository and produces structured documentation that explains the project structure, key components, architectural relationships, and intended system behavior. The blueprint leverages LLMs to reduce manual documentation efforts, while improving code understanding and accelerating developer onboarding.

AMD Solution Blueprints are packaged as Helm charts for deployment on a Kubernetes cluster. For development or further exploration, the source code is public and available in the Solution Blueprints GitHub repository.

Architecture#

Code Docs Builder uses three key components: AIM LLM (Llama 3.3 70B Instruct model), backend, and frontend. Users interact with the frontend, which calls the backend to clone the repository and start the documentation generation process using CrewAI tools to instruct the LLM how to act and produce the final documentation.

The blueprint deploys Code Docs Builder as a containerized web application with pre-configured crews (that combine agents and tasks) and integrated LLM connectivity through AIMs for seamless AI agent orchestration.

Component	Role
AIM LLM	Large Language Model for powering the agents; default in this blueprint is Llama-3.3-70B
CrewAI	Preconfigured agents and tasks that orchestrate repository analysis and LLM-driven documentation generation
Python/FastAPI	A backend service that exposes APIs and handles orchestration for the web interface
Python/Gradio	A lightweight web interface for interacting with ML models and agent workflows
Kubernetes	Container orchestration and deployment platform

⚠️ Important: This blueprint has been tested and validated with the AIM using Llama 3.3 70B Instruct. While other models may work, their compatibility and performance are not guaranteed.

⚠️ Warning: This blueprint is created for demonstration purposes. It has some limitations when the repository size is large, and processing may take a while (sometimes hours). In such cases, the quality of the generated documentation may be worse than for smaller repositories. Adjust the agent prompts in config/ to better suit your use case and repository structure.

Key Features#

Automatic Codebase Analysis: Agents automatically analyze the codebase and generate diagrams that reflect the project’s structure and architecture
Structured Documentation Output: Documentation is planned and generated in well-defined sections, each serving a specific purpose
Zero Configuration Required: Documentation generation relies on preconfigured agents and tasks — no manual setup needed
Simple and Intuitive Workflow: Provide a GitHub repository link and generate documentation in a few clicks
Ideal for Developer Onboarding: Onboard new developers into projects with little to no existing documentation
Downloadable Results: Export the final documentation in clean, ready-to-use .md format

Getting Started#

This is a quick start guide on how to deploy the blueprint. For advanced options, such as reusing an existing AIM, providing a Hugging Face token, or overriding storage classes, see Deploying Solution Blueprints with Helm or explore the advanced deployment guide.

Prerequisites#

System Requirements#

This blueprint can be deployed on AMD Instinct (default) and AMD EPYC. The blueprint requires the following cluster resources by default, depending on the hardware being used:

Resource	Instinct	EPYC
GPUs	1	—
CPUs	4 CPU cores	188 CPU cores
RAM	64 Gi	128 Gi

To deploy to the Kubernetes cluster, ensure the following prerequisites are met:

kubectl: Installed and configured to communicate with the cluster
Helm 3.17 or higher: Installed on your local machine

Deployment#

For advanced deployment options, explore the advanced deployment guide. Solution Blueprints are packaged as OCI-compliant Helm charts in the Docker Hub registry and can be deployed to a Kubernetes cluster with a single command. Define the name (deployment name) and the namespace (Kubernetes namespace), then pipe the output of helm template to kubectl apply -f -.

Find the deployment command below. Note: You can create a namespace using kubectl create namespace <my-namespace>.

Instinct

name="my-deployment"
namespace="my-namespace"
helm template $name oci://registry-1.docker.io/amdenterpriseai/aimsb-codedocs \
  | kubectl apply -f - -n $namespace

EPYC

EPYC runs the model on CPU (gpus=0, bf16, AIM_ALLOW_UNOPTIMIZED=true), sized via llm.cpus/llm.memory. The default EPYC AIM is a gated image, so provide a Hugging Face token through a Secret.

name="my-deployment"
namespace="my-namespace"
kubectl create namespace $namespace
kubectl create secret generic hf-token --from-literal=hf-token=<YOUR_HF_TOKEN> -n $namespace

helm pull oci://registry-1.docker.io/amdenterpriseai/aimsb-codedocs --untar
helm template $name ./aimsb-codedocs \
  --set global.platform=epyc \
  --set llm.cpus=188 \
  --set llm.memory=128 \
  --set llm.env_vars.HF_TOKEN.name=hf-token \
  --set llm.env_vars.HF_TOKEN.key=hf-token \
  | kubectl apply -f - -n $namespace

Performance note: On multi-socket EPYC nodes, configure the kubelet for NUMA alignment (CPU Manager static, Topology Manager single-numa-node, Memory Manager Static); otherwise the LLM’s CPUs and memory can land on different NUMA nodes and vLLM runs effectively single-threaded.

Verify Deployment#

To check the status of the deployment, run:

kubectl get pods -n $namespace

Wait until all pods report Running and Ready. Code documentation works only when the LLM service is also up and running. Note that the default model used is large and can take ~10 minutes to start.

Connect to UI#

To connect to the UI, port-forward to 8092. The UI will then be available at http://localhost:8092 in your browser.

kubectl port-forward services/${name}-aimsb-codedocs-frontend 8092:8092 -n $namespace

Once connected, use the application as follows:

Provide a GitHub repository link and generate documentation in a few clicks
Export the final documentation in clean, ready-to-use .md format

Clean Up#

When you are finished, remove the deployed resources using the same deployment command, with kubectl delete instead of kubectl apply. For example, for Instinct use the following command:

helm template $name oci://registry-1.docker.io/amdenterpriseai/aimsb-codedocs \
  | kubectl delete -f - -n $namespace

Third-Party Components#

This Solution Blueprint uses multiple third-party components. To see the full set of software and Python dependencies, explore the repository source and dependency files. The table below highlights some of the key components. For further license information, refer to each component’s official documentation.

Component	License
CrewAI	MIT
FastAPI	MIT
Gradio	Apache 2.0

Terms of Use#

AMD Solution Blueprints are released under the MIT License, which governs the parts of the software and materials created by AMD. Third-party Software and Materials used within the Solution Blueprints are governed by their respective licenses.