Skip to content

README.md

Latest commit

 

History

History
62 lines (43 loc) · 2.85 KB

README.md

File metadata and controls

62 lines (43 loc) · 2.85 KB
Raw

Galileo LLM-as-Judge Direct Evaluator Example

This example shows an Agent Control agent using the direct Galileo LLM-as-judge evaluator (galileo.llm). The evaluator calls the same scorer invoke URL as the Luna evaluator and applies thresholds locally from the control definition. It is suited for boolean scorers (e.g. prompt injection detection) where the scorer returns True/False.

What It Shows

  • setup_controls.py registers an agent and attaches three controls.
  • demo_agent.py runs an agent step protected with @control.
  • A pre-execution control scores the raw user input with galileo.llm and blocks before the LLM runs if the scorer flags it as a prompt injection attempt.
  • A composite post-execution control combines a built-in list evaluator (fast keyword pre-filter) with galileo.llm scoring the LLM output.
  • A regex control blocks leaked API-key-like values in generated output.

Setup

Start the Agent Control server from the repo root:

make server-run

Configure LLM scorer invoke credentials:

export GALILEO_API_SECRET_KEY="your-api-secret"
export GALILEO_LUNA_INVOKE_URL="http://luna-invoke.internal/api/v1/scorers/invoke"

GALILEO_API_SECRET can be used instead of GALILEO_API_SECRET_KEY if that is how your deployment exposes the internal Galileo JWT signing secret. GALILEO_LUNA_INVOKE_URL is shared between the Luna and LLM evaluators — both use the same scorer invoke endpoint.

Required scorer setting:

export GALILEO_LLM_SCORER_ID="your-scorer-uuid"

Optional scorer settings:

export GALILEO_LLM_SCORER_LABEL="prompt_injection_llm"   # display/metadata label only
export GALILEO_LLM_SCORER_VERSION_ID="version-uuid"      # optional compatibility ID
export GALILEO_LLM_OPERATOR="eq"                         # default: eq (for boolean scorers)
export GALILEO_LLM_PAYLOAD_FIELD="input"                 # default: input

GALILEO_LLM_OPERATOR and the threshold (True by default) work together. For a boolean prompt injection scorer that returns True when injection is detected, operator=eq and threshold=True means: block when the scorer returns True.

GALILEO_LLM_PAYLOAD_FIELD controls which scorer input side is used when the selector returns a scalar value. The pre-execution control selects input (the raw user message) and the post-execution composite control selects output (the LLM reply). When a selector returns structured data with input and/or output keys, those keys are sent directly and override this setting.

If the invoke endpoint uses an internal certificate authority, configure one of:

export GALILEO_LUNA_INVOKE_CA_FILE="/etc/ssl/internal/luna-invoke-ca.crt"
export AGENT_CONTROL_AUTH_UPSTREAM_CA_FILE="/etc/agent-control/auth-upstream-ca/ca.crt"

Run:

cd examples/galileo_llm
uv run python setup_controls.py
uv run python demo_agent.py