Fair Forge Lambda

Fair-Forge Lambda Deployment

Generate AWS Lambda container deployment for Fair-Forge modules. Builds wheel from local repository since fair-forge is not available on PyPI.

Supported Modules

Workflow

1. Generate Lambda Project

Create the deployment directory:

examples/{module}/aws-lambda/
├── Dockerfile       # Multi-stage build: wheel + Lambda
├── handler.py       # Lambda entrypoint
├── run.py           # Business logic (customize this)
├── requirements.txt # Additional runtime dependencies
├── README.md        # API documentation
└── scripts/
    ├── deploy.sh
    ├── update.sh
    └── cleanup.sh

Copy templates from assets/templates/ and scripts from scripts/.

2. CRITICAL: Update Dockerfile Paths

The Dockerfile is built with the repo root as context (not the aws-lambda directory). You MUST update the COPY paths to use full paths from the repo root:

# UPDATE these lines in the Dockerfile:
COPY examples/{module}/aws-lambda/requirements.txt .
COPY examples/{module}/aws-lambda/handler.py examples/{module}/aws-lambda/run.py ${LAMBDA_TASK_ROOT}/

For example, for generators:

COPY examples/generators/aws-lambda/requirements.txt .
COPY examples/generators/aws-lambda/handler.py examples/generators/aws-lambda/run.py ${LAMBDA_TASK_ROOT}/

3. Implement run.py

Edit run.py with your module logic. See reference docs for patterns:

• references/metrics.md - Metrics implementation patterns
• references/runners.md - Runners implementation patterns
• references/generators.md - Generators implementation patterns

4. Generate README.md

After deployment, create a README.md using the template from assets/templates/README.md. Replace the placeholders:

See the template for full list of placeholders.

5. Deploy

From examples/{module}/aws-lambda/:

./scripts/deploy.sh {extra-name} us-east-2

Examples:

./scripts/deploy.sh bestof us-east-2      # BestOf metric
./scripts/deploy.sh runners us-east-2     # Runners module
./scripts/deploy.sh generators us-east-2  # Generators module

6. Update / Cleanup

./scripts/update.sh {extra-name} us-east-2   # Rebuild and update
./scripts/cleanup.sh {extra-name} us-east-2  # Remove all resources

Dynamic LLM Connector API

For modules that use LLMs (generators, metrics), the API uses a dynamic connector pattern that allows runtime selection of LLM provider:

Request Format

{
  "connector": {
    "class_path": "langchain_groq.chat_models.ChatGroq",
    "params": {
      "model": "qwen/qwen3-32b",
      "api_key": "your-api-key",
      "temperature": 0.7
    }
  },
  "context": "...",
  "config": {...}
}

Supported Connectors

Connector Factory Implementation

The create_llm_connector function uses dynamic imports to instantiate any LangChain chat model:

import importlib

def create_llm_connector(connector_config: dict):
    class_path = connector_config.get("class_path")
    params = connector_config.get("params", {})

    module_path, class_name = class_path.rsplit(".", 1)
    module = importlib.import_module(module_path)
    cls = getattr(module, class_name)

    return cls(**params)

Dockerfile Build Process

Multi-stage build:

1. Builder stage: Builds fair-forge wheel from local source
2. Final stage: Pre-installs numpy/scipy, installs wheel with module extras, copies handler

numpy/scipy Compilation Workaround

The Lambda base image lacks compilers needed to build numpy/scipy from source. The Dockerfile pre-installs specific versions with pre-built manylinux wheels:

# Pre-install numpy and scipy with manylinux wheels
RUN pip install "numpy==1.26.4" "scipy==1.14.1" --target "${LAMBDA_TASK_ROOT}"

# Constraints file prevents pip from upgrading these
RUN echo "numpy==1.26.4" > /tmp/constraints.txt && \
    echo "scipy==1.14.1" >> /tmp/constraints.txt

# Install fair-forge with constraints
ARG MODULE_EXTRA=bestof
RUN WHEEL=$(ls /tmp/*.whl) && pip install "${WHEEL}[${MODULE_EXTRA}]" --target "${LAMBDA_TASK_ROOT}" -c /tmp/constraints.txt

Bundled Resources

scripts/

• deploy.sh - Full deployment (ECR + Lambda + API Gateway)
• update.sh - Rebuild and update Lambda
• cleanup.sh - Remove all AWS resources

assets/templates/

• Dockerfile - Multi-stage Lambda container build
• handler.py - Lambda entrypoint wrapper
• run.py - Business logic template (customize this)
• requirements.txt - Runtime dependencies with all LLM providers
• README.md - API documentation template with placeholders

references/

• metrics.md - Metrics module implementation patterns
• runners.md - Runners module implementation patterns
• generators.md - Generators module implementation patterns

Lambda Configuration

Default settings (adjust in deploy.sh):

• Timeout: 300 seconds (5 minutes)
• Memory: 2048 MB
• Python: 3.11

Troubleshooting

"file does not exist" during Docker build

The Dockerfile paths for requirements.txt, handler.py, and run.py must be updated to use full paths from the repo root. See step 2 above.

scipy/numpy compilation errors

The template Dockerfile includes a workaround that pre-installs numpy and scipy from manylinux wheels. If you see errors like "NumPy requires GCC >= 9.3", ensure you're using the updated Dockerfile template.

Wheel path duplication error

If you see /tmp//tmp/filename.whl, the wheel installation command is incorrect. Use:

RUN WHEEL=$(ls /tmp/*.whl) && pip install "${WHEEL}[${MODULE_EXTRA}]" ...

NOT:

RUN pip install "/tmp/$(ls /tmp/*.whl)[${MODULE_EXTRA}]" ...

langchain import errors (e.g., "cannot import name 'ModelProfile'")

This indicates a version mismatch between langchain packages. Pin compatible versions in requirements.txt:

langchain-core>=0.3.0,<0.4.0
langchain-groq>=0.2.0,<0.3.0

tiktoken Rust compiler error

If you see "can't find Rust compiler" when installing langchain-openai, pin tiktoken to a version with pre-built wheels:

tiktoken>=0.7.0,<0.8.0