prellm 0.1.15

Prellm - Lightweight LLM prompt middleware for bias detection, standardization, and DevOps process chains via YAML config, supporting context injection and multi-provider orchestration.

🛡️ Prellm

Lightweight LLM prompt middleware — bias detection, standardization, and DevOps process chains via YAML config.

Prellm sits between your application and LLM providers, automatically detecting bias, ambiguity, and dangerous patterns in prompts. It enriches queries with context, validates outputs, and supports multi-step DevOps workflows with approval gates.

Features

• Bias & Ambiguity Detection — regex + NLTK patterns for PL/EN, with DevOps-specific guardrails
• YAML-Driven Config — declarative rules, clarification templates, model fallbacks
• 100+ LLM Models — via LiteLLM proxy (OpenAI, Anthropic, Llama, Mistral, etc.)
• DevOps Process Chains — multi-step workflows with approval gates, rollback, and audit trails
• Context Injection — auto-enrich prompts with env vars, git info, system state
• Type-Safe Outputs — Pydantic v2 validated responses
• Lightweight — <50MB, 5 dependencies, async-first

Quick Start

# Install
pip install prellm

# Generate config
prellm init --devops -o rules.yaml

# Analyze a query (no LLM call)
prellm analyze "Deploy to production" --config rules.yaml

# Run with LLM
prellm run "Zdeployuj na staging" --config rules.yaml --model gpt-4o-mini

# Execute a process chain
prellm process deploy.yaml --guard-config rules.yaml --env production

Python API

from prellm import preLLM, ProcessChain

# Simple query
guard = preLLM("rules.yaml")
result = await guard("Deploy to production", model="gpt-4o-mini")
print(result.clarified)  # True — detected missing context
print(result.content)     # Enriched response

# Process chain
chain = ProcessChain("deploy.yaml")
result = await chain.execute(env="production", dry_run=True)
for step in result.steps:
    print(f"{step.step_name}: {step.status}")

Configuration

rules.yaml

bias_patterns:
  - regex: "(deploy|zdeployuj)\\s+(na|to)\\s+(prod|production)"
    action: clarify
    severity: critical
    description: "Production deployment — requires context"

clarify_template: "[KONTEKST]: Podaj szczegóły dla: {query}"
max_retries: 3
policy: devops

models:
  fallback: ["gpt-4o-mini", "llama3"]

context_sources:
  - env: [CLUSTER, NAMESPACE, GIT_SHA]
  - git: [branch, short_sha]

deploy.yaml (Process Chain)

process: deploy-production
steps:
  - name: pre-check
    prompt: "Check readiness of {CLUSTER}"
    approval: auto
  - name: deploy
    prompt: "Rolling deploy to {CLUSTER}/{NAMESPACE}"
    approval: manual
    rollback: true

Architecture

User Query → BiasDetector → ContextEngine → Enrichment → LiteLLM → Pydantic Validation → Response
                                                              ↑
                                        ProcessChain → Approval Gates → Audit Trail

Development

git clone https://github.com/softreck/prellm
cd prellm
poetry install
poetry run pytest

License

Apache License 2.0 - see LICENSE for details.

Author

Created by Tom Sapletta - [e-mail]