haiku.rag/haiku_rag_slim/haiku/rag/agents/analysis/agent.py
2026-04-17 18:35:16 +03:00

59 lines
1.9 KiB
Python

from pydantic_ai import Agent, RunContext
from haiku.rag.agents.analysis.dependencies import AnalysisDeps
from haiku.rag.agents.analysis.models import CodeExecution, RawAnalysisResult
from haiku.rag.agents.analysis.prompts import ANALYSIS_SYSTEM_PROMPT
from haiku.rag.config.models import AppConfig
from haiku.rag.utils import get_model
def create_analysis_agent(config: AppConfig) -> Agent[AnalysisDeps, RawAnalysisResult]:
"""Create an analysis agent with code execution capability.
The analysis agent can write and execute Python code in a sandboxed
environment to solve problems that require computation, aggregation,
or complex traversal across documents.
Args:
config: Application configuration.
Returns:
A pydantic-ai Agent configured for analysis execution.
"""
model = get_model(config.analysis.model, config)
agent: Agent[AnalysisDeps, RawAnalysisResult] = Agent( # type: ignore[assignment] # ty: ignore[invalid-assignment]
model,
deps_type=AnalysisDeps,
output_type=RawAnalysisResult,
instructions=ANALYSIS_SYSTEM_PROMPT,
retries=3,
)
@agent.tool
async def execute_code(ctx: RunContext[AnalysisDeps], code: str) -> CodeExecution:
"""Execute Python code in a sandboxed interpreter.
The code has access to search() and llm() functions, and a
virtual filesystem at /documents/ with document content and structure.
Use print() to output results.
Args:
code: Python code to execute.
Returns:
Structured result with success status, stdout, and stderr.
"""
result = await ctx.deps.sandbox.execute(code)
execution = CodeExecution(
code=code,
stdout=result.stdout,
stderr=result.stderr,
success=result.success,
)
return execution
return agent