BLUEPATTERNSAI · AWARENESS COOKBOOK
Ground and evaluate any LLM response.
Place Awareness between your application and its foundational model. Requests are routed, grounded in certified context, evaluated, and returned with citations—or blocked safely.
01 · Overview & use case
Build a governed LLM call
Send questions to the OpenAI-compatible Awareness endpoint instead of calling the model directly. Awareness retrieves approved legal evidence, evaluates grounding and safety, and exposes citations and token telemetry.
sarvam-105b model. Other providers or domains require certification.02 · Prerequisites & setup
Install nothing beyond Python
python3 --version
export AWARENESS_API_KEY='<tenant-key-from-secure-onboarding>'
export AWARENESS_BASE_URL='https://awareness.bluepatterns.ai/v1'
python3 awareness_cookbook.pyNever put admin or provider credentials in browser code, source control, or a client bundle.
03 · Core implementation
Runnable Python client
Save awareness_cookbook.py beside this page and run it. It handles timeouts and malformed responses without bypassing evaluation.
from awareness_cookbook import call_awareness
result = call_awareness("What is the exact authorization protocol for a $6,000 software purchase under Section 99-B?")
print(result.output, result.citations, result.usage)The result contains the evaluated answer, source citations, usage counters, and cache state.
04 · Good vs. bad patterns
Keep the evaluation boundary intact
| 🚫 Suboptimal approach | Optimized pattern |
|---|---|
Call the provider from the UI.fetch(PROVIDER_URL) | Route every request through Awareness.base_url="https://awareness.bluepatterns.ai/v1" |
| Trust any HTTP 200 response. | Require the decision, citations, and usage fields. |
| Put keys in React/browser code. | Keep tenant keys in a server-side secret manager. |
| Retry by dropping evidence or safety checks. | Let Awareness repack, retry, or fail closed. |
05 · Production considerations
Operate the boundary deliberately
- Use one tenant key per client environment; rotate through onboarding and never log it.
- Respect context and output budgets; Awareness enforces the configured window.
- Treat timeouts, quota errors, malformed JSON, and failed evaluation as blocked outcomes.
- Retain citations and usage telemetry for audit; redact prompts, PII, headers, and rejected outputs.