AI Integration

Lexigram provides a comprehensive suite of AI primitives designed for security, observability, and modularity. The lexigram-ai platform handles everything from multi-provider LLM orchestration to complex RAG pipelines and autonomous agents.

---

1. Minimal LLM Setup

The core of the platform is the LLMClientProtocol. You can switch between 18+ providers (OpenAI, Anthropic, Gemini, etc.) using only configuration.

Configuration

application.yaml

ai:
  llm:
    provider: anthropic
    model: claude-sonnet-4-6
    temperature: 0.7

Usage

from lexigram.contracts.ai.llm import LLMClientProtocol
from lexigram.contracts.ai.types import ChatMessage, Role

class ChatService:
    def __init__(self, llm: LLMClientProtocol) -> None:
        self.llm = llm

    async def get_response(self, prompt: str) -> str:
        response = await self.llm.complete(
            messages=[ChatMessage(role=Role.USER, content=prompt)]
        )
        return response.content

---

2. RAG Pipelines (Retrieval Augmented Generation)

Lexigram RAG coordinates document loading, embedding, vector storage, and retrieval.

from lexigram.ai.rag import RAGPipelineProtocol

async def ask_documents(rag: RAGPipelineProtocol, query: str):
    # This single call metadata retrieval, context synthesis,
    # and LLM generation.
    result = await rag.query(query)

    print(f"Answer: {result.answer}")
    print(f"Sources: {result.sources}")

---

3. Autonomous Agents

Agents in Lexigram can use Tools, maintain Memory, and follow orchestrations like ReAct or Plan-and-Execute.

Defining a Tool

from lexigram.ai.agents import tool

@tool
async def get_stock_price(ticker: str) -> float:
    """Retrieves the real-time stock price for a given ticker symbol."""
    return 150.0  # Integration logic here

Executing an Agent

from lexigram.ai.agents import AgentExecutorProtocol

async def run_investor_agent(executor: AgentExecutorProtocol):
    response = await executor.run(
        "Should I buy AAPL? Check its current price first.",
        tools=["get_stock_price"]
    )
    print(response.output)

---

4. Multi-Provider Routing & Resilience

Lexigram can automatically route requests through multiple providers to balance cost, latency, or availability.

application.yaml

ai:
  llm:
    routing:
      strategy: sequential
      providers:
        - name: anthropic
          model: claude-sonnet-4-6
        - name: openai
          model: gpt-4o         # Fallback if Anthropic fails

---

5. Governance & Observability

AI workflows are notoriously opaque. Lexigram provides built-in tools to monitor and control them.

• Tracing: Full visibility into agent reasoning loops and RAG retrieval steps.
• Guards: Input/Output filtering to prevent prompt injection or PII leakage.
• Cost Tracking: Real-time token usage metrics and cost estimation per model.

---

[!TIP] Use Thinking Suppression for models like Claude Sonnet or OpenAI o1 to hide internal reasoning tokens while still benefiting from their advanced logic. This is configured in ThinkingConfig at the provider level.

---

Technical Stack

Package	Purpose
lexigram-ai-llm	Multi-provider client & routing.
lexigram-ai-rag	Document ingestion & retrieval.
lexigram-ai-agents	ReAct, Planning, and Tool use.
lexigram-ai-memory	Short & long-term AI memory.
lexigram-ai-governance	AI policy and audit logs.
lexigram-ai-guard	Input/Output security filters.