masterllm / services /agent_crewai.py
redhairedshanks1's picture
Update services/agent_crewai.py
dc7a501
# services/agent_crewai.py
"""
CrewAI-based agent for MasterLLM orchestration.
"""
import json
import os
from typing import Optional, Dict, Any, List, Generator
from crewai import Agent, Task, Crew, Process
from crewai.tools import BaseTool
from pydantic import BaseModel, Field
# Import your remote utilities
from utilities.extract_text import extract_text_remote
from utilities.extract_tables import extract_tables_remote
from utilities.describe_images import describe_images_remote
from utilities.summarizer import summarize_remote
from utilities.classify import classify_remote
from utilities.ner import ner_remote
from utilities.translator import translate_remote
from utilities.signature_verification import signature_verification_remote
from utilities.stamp_detection import stamp_detection_remote
# ========================
# TOOL INPUT SCHEMAS
# ========================
class FileSpanInput(BaseModel):
file_path: str = Field(..., description="Absolute/local path to the uploaded file")
start_page: int = Field(1, description="Start page (1-indexed)")
end_page: int = Field(1, description="End page (inclusive, 1-indexed)")
class TextOrFileInput(BaseModel):
text: Optional[str] = Field(None, description="Raw text to process")
file_path: Optional[str] = Field(None, description="Path to a document on disk (PDF/Image)")
start_page: int = Field(1, description="Start page (1-indexed)")
end_page: int = Field(1, description="End page (inclusive, 1-indexed)")
class TranslateInput(TextOrFileInput):
target_lang: str = Field(..., description="Target language code or name (e.g., 'es' or 'Spanish')")
# ========================
# HELPER FUNCTIONS
# ========================
def _base_state(file_path: str, start_page: int = 1, end_page: int = 1) -> Dict[str, Any]:
"""Build the base state your utilities expect."""
filename = os.path.basename(file_path)
return {
"filename": filename,
"temp_files": {filename: file_path},
"start_page": start_page,
"end_page": end_page,
}
# ========================
# CREWAI TOOLS
# ========================
class ExtractTextTool(BaseTool):
name: str = "extract_text"
description: str = """Extract text from a document between start_page and end_page (inclusive).
Use this when the user asks to read, analyze, or summarize document text.
Input should be a JSON object with: file_path (required), start_page (default 1), end_page (default 1)."""
def _run(self, file_path: str, start_page: int = 1, end_page: int = 1) -> str:
state = _base_state(file_path, start_page, end_page)
out = extract_text_remote(state)
text = out.get("text") or out.get("extracted_text") or ""
return json.dumps({"text": text})
class ExtractTablesTool(BaseTool):
name: str = "extract_tables"
description: str = """Extract tables from a document between start_page and end_page.
Input should be a JSON object with: file_path (required), start_page (default 1), end_page (default 1)."""
def _run(self, file_path: str, start_page: int = 1, end_page: int = 1) -> str:
state = _base_state(file_path, start_page, end_page)
out = extract_tables_remote(state)
tables = out.get("tables", [])
return json.dumps({"tables": tables, "table_count": len(tables)})
class DescribeImagesTool(BaseTool):
name: str = "describe_images"
description: str = """Generate captions/descriptions for images in the specified page range.
Input should be a JSON object with: file_path (required), start_page (default 1), end_page (default 1)."""
def _run(self, file_path: str, start_page: int = 1, end_page: int = 1) -> str:
state = _base_state(file_path, start_page, end_page)
out = describe_images_remote(state)
return json.dumps({"image_descriptions": out.get("image_descriptions", out)})
class SummarizeTextTool(BaseTool):
name: str = "summarize_text"
description: str = """Summarize either raw text or a document (by file_path + optional page span).
Input should be a JSON object with: text (optional), file_path (optional), start_page (default 1), end_page (default 1).
At least one of text or file_path must be provided."""
def _run(
self,
text: Optional[str] = None,
file_path: Optional[str] = None,
start_page: int = 1,
end_page: int = 1,
) -> str:
state: Dict[str, Any] = {
"text": text,
"start_page": start_page,
"end_page": end_page,
}
if file_path:
state.update(_base_state(file_path, start_page, end_page))
out = summarize_remote(state)
return json.dumps({"summary": out.get("summary", out)})
class ClassifyTextTool(BaseTool):
name: str = "classify_text"
description: str = """Classify a text or document content.
Input should be a JSON object with: text (optional), file_path (optional), start_page (default 1), end_page (default 1).
At least one of text or file_path must be provided."""
def _run(
self,
text: Optional[str] = None,
file_path: Optional[str] = None,
start_page: int = 1,
end_page: int = 1,
) -> str:
state: Dict[str, Any] = {
"text": text,
"start_page": start_page,
"end_page": end_page,
}
if file_path:
state.update(_base_state(file_path, start_page, end_page))
out = classify_remote(state)
return json.dumps({"classification": out.get("classification", out)})
class ExtractEntitesTool(BaseTool):
name: str = "extract_entities"
description: str = """Perform Named Entity Recognition (NER) on text or a document.
Input should be a JSON object with: text (optional), file_path (optional), start_page (default 1), end_page (default 1).
At least one of text or file_path must be provided."""
def _run(
self,
text: Optional[str] = None,
file_path: Optional[str] = None,
start_page: int = 1,
end_page: int = 1,
) -> str:
state: Dict[str, Any] = {
"text": text,
"start_page": start_page,
"end_page": end_page,
}
if file_path:
state.update(_base_state(file_path, start_page, end_page))
out = ner_remote(state)
return json.dumps({"ner": out.get("ner", out)})
class TranslateTextTool(BaseTool):
name: str = "translate_text"
description: str = """Translate text or a document to target_lang (e.g., 'es', 'fr', 'de', 'Spanish').
Input should be a JSON object with: target_lang (required), text (optional), file_path (optional),
start_page (default 1), end_page (default 1). At least one of text or file_path must be provided."""
def _run(
self,
target_lang: str,
text: Optional[str] = None,
file_path: Optional[str] = None,
start_page: int = 1,
end_page: int = 1,
) -> str:
state: Dict[str, Any] = {
"text": text,
"start_page": start_page,
"end_page": end_page,
"target_lang": target_lang,
}
if file_path:
state.update(_base_state(file_path, start_page, end_page))
out = translate_remote(state)
return json.dumps({
"translation": out.get("translation", out),
"target_lang": target_lang
})
class SignatureVerificationTool(BaseTool):
name: str = "signature_verification"
description: str = """Verify signatures/stamps presence and authenticity indicators in specified page range.
Input should be a JSON object with: file_path (required), start_page (default 1), end_page (default 1)."""
def _run(self, file_path: str, start_page: int = 1, end_page: int = 1) -> str:
state = _base_state(file_path, start_page, end_page)
out = signature_verification_remote(state)
return json.dumps({"signature_verification": out.get("signature_verification", out)})
class StampDetectionTool(BaseTool):
name: str = "stamp_detection"
description: str = """Detect stamps in a document in the specified page range.
Input should be a JSON object with: file_path (required), start_page (default 1), end_page (default 1)."""
def _run(self, file_path: str, start_page: int = 1, end_page: int = 1) -> str:
state = _base_state(file_path, start_page, end_page)
out = stamp_detection_remote(state)
return json.dumps({"stamp_detection": out.get("stamp_detection", out)})
# ========================
# TOOL REGISTRY
# ========================
def get_master_tools() -> List[BaseTool]:
"""Export all tools for CrewAI agent binding."""
return [
ExtractTextTool(),
ExtractTablesTool(),
DescribeImagesTool(),
SummarizeTextTool(),
ClassifyTextTool(),
ExtractEntitesTool(),
TranslateTextTool(),
SignatureVerificationTool(),
StampDetectionTool(),
]
# ========================
# AGENT CONFIGURATION
# ========================
SYSTEM_INSTRUCTIONS = """You are MasterLLM, a precise document processing agent.
Your responsibilities:
- Use tools for any action (extraction, tables, images, summarization, classification, NER, translation, signature verification, stamp detection).
- If a tool requires file_path and the user didn't provide one, use the provided session_file_path.
- Use page spans when relevant (start_page, end_page).
- Combine results when needed (e.g., extract_text -> summarize_text; tables -> summarize_text).
- If a PLAN is provided, follow it strictly unless it's impossible.
- Keep outputs compact - do not include raw base64 or giant blobs.
- Always return a final JSON result with:
{
"steps_executed": [...],
"outputs": { ... },
"errors": [],
"meta": {
"model": "crewai-gemini",
"notes": "short note if needed"
}
}
"""
def create_master_agent(session_file_path: str = "", plan_json: str = "{}") -> Agent:
"""Create the master document processing agent."""
tools = get_master_tools()
backstory = f"""{SYSTEM_INSTRUCTIONS}
Current session file: {session_file_path}
Execution plan: {plan_json}
"""
# Use LiteLLM-compatible string format for Gemini
# CrewAI internally uses LiteLLM which requires "gemini/" prefix
# Now that google-generativeai is installed, this should work
try:
import google.generativeai as genai
# Verify package is installed and configure it
api_key = os.getenv("GOOGLE_API_KEY") or os.getenv("GEMINI_API_KEY")
if api_key:
genai.configure(api_key=api_key)
# Use LiteLLM string format (not the model object)
llm = "gemini/gemini-2.0-flash-exp"
print(f"✅ CrewAI using Gemini via LiteLLM: {llm}")
else:
print("⚠️ No Google API key found")
llm = "gemini/gemini-2.0-flash-exp"
except ImportError as e:
print(f"⚠️ google.generativeai not available: {e}")
print("Using LiteLLM format anyway")
llm = "gemini/gemini-2.0-flash-exp"
except Exception as e:
print(f"⚠️ Error configuring Gemini: {e}")
llm = "gemini/gemini-2.0-flash-exp"
agent = Agent(
role="Document Processing Specialist",
goal="Process documents according to the given plan using available tools, and return structured JSON results",
backstory=backstory,
tools=tools,
verbose=True,
allow_delegation=False,
max_iter=12,
llm=llm,
)
return agent
def create_master_crew(
user_input: str,
session_file_path: str = "",
plan: Optional[Dict[str, Any]] = None,
) -> Crew:
"""Create a crew with the master agent and a task based on user input."""
plan_json = json.dumps(plan or {})
agent = create_master_agent(session_file_path, plan_json)
task_description = f"""
Execute the following document processing request:
User Request: {user_input}
Session File Path: {session_file_path}
Execution Plan: {plan_json}
Instructions:
1. Follow the plan steps in order
2. Use the file path provided for all file-based operations
3. Combine results from multiple tools when appropriate
4. Return a comprehensive JSON result with all outputs
Expected Output Format:
{{
"steps_executed": ["step1", "step2", ...],
"outputs": {{
"text": "...",
"tables": [...],
"summary": "...",
// other outputs based on what was executed
}},
"errors": [],
"meta": {{
"model": "crewai-gemini",
"pipeline": "{plan.get('pipeline', '') if plan else ''}",
"pages_processed": "{plan.get('start_page', 1)}-{plan.get('end_page', 1) if plan else '1-1'}"
}}
}}
"""
task = Task(
description=task_description,
expected_output="A JSON object containing all processed results, executed steps, and any errors",
agent=agent,
)
crew = Crew(
agents=[agent],
tasks=[task],
process=Process.sequential,
verbose=True,
)
return crew
# ========================
# MAIN ENTRY POINTS
# ========================
def run_agent(
user_input: str,
session_file_path: Optional[str] = None,
plan: Optional[Dict[str, Any]] = None,
chat_history: Optional[List[Any]] = None,
) -> Dict[str, Any]:
"""
Invokes the CrewAI agent to process the document.
Returns a dict with the processing results.
"""
crew = create_master_crew(
user_input=user_input,
session_file_path=session_file_path or "",
plan=plan,
)
result = crew.kickoff()
# Parse the result - CrewAI returns a CrewOutput object
try:
if hasattr(result, 'raw'):
raw_output = result.raw
else:
raw_output = str(result)
# Try to parse as JSON
try:
parsed = json.loads(raw_output)
return {"output": parsed}
except json.JSONDecodeError:
# Try to extract JSON from the response
import re
json_match = re.search(r'\{.*\}', raw_output, re.DOTALL)
if json_match:
try:
parsed = json.loads(json_match.group())
return {"output": parsed}
except json.JSONDecodeError:
pass
# Return as-is if not JSON
return {"output": {"result": raw_output, "format": "text"}}
except Exception as e:
return {"output": {"error": str(e), "raw_result": str(result)}}
def run_agent_streaming(
user_input: str,
session_file_path: Optional[str] = None,
plan: Optional[Dict[str, Any]] = None,
chat_history: Optional[List[Any]] = None,
) -> Generator[Dict[str, Any], None, None]:
"""
Streaming version of run_agent that yields intermediate step updates.
Each yield contains: {"type": "step"|"final", "data": {...}}
Note: CrewAI doesn't have native streaming like LangChain's AgentExecutor,
so we simulate it by yielding progress updates and then the final result.
"""
import threading
import queue
import time
result_queue: queue.Queue = queue.Queue()
# Yield initial status
yield {
"type": "step",
"step": 0,
"status": "initializing",
"tool": "crew_setup",
"input_preview": f"Setting up pipeline: {plan.get('pipeline', 'unknown') if plan else 'unknown'}"
}
def run_crew():
try:
crew = create_master_crew(
user_input=user_input,
session_file_path=session_file_path or "",
plan=plan,
)
result = crew.kickoff()
result_queue.put(("success", result))
except Exception as e:
result_queue.put(("error", str(e)))
# Start crew execution in a separate thread
thread = threading.Thread(target=run_crew)
thread.start()
# Yield progress updates while waiting
step_count = 1
pipeline_steps = plan.get("pipeline", "").split("-") if plan else []
for step_name in pipeline_steps:
yield {
"type": "step",
"step": step_count,
"status": "executing",
"tool": step_name,
"input_preview": f"Processing: {step_name}"
}
step_count += 1
# Check if result is ready
try:
result_type, result_data = result_queue.get(timeout=2.0)
break
except queue.Empty:
continue
# Wait for completion if not already done
thread.join(timeout=120) # Max 2 minutes timeout
# Get final result
try:
if result_queue.empty():
yield {
"type": "error",
"error": "Execution timeout - crew did not complete in time"
}
return
result_type, result_data = result_queue.get_nowait()
if result_type == "error":
yield {
"type": "error",
"error": result_data
}
return
# Parse the result
try:
if hasattr(result_data, 'raw'):
raw_output = result_data.raw
else:
raw_output = str(result_data)
# Try to parse as JSON
try:
parsed = json.loads(raw_output)
except json.JSONDecodeError:
import re
json_match = re.search(r'\{.*\}', raw_output, re.DOTALL)
if json_match:
try:
parsed = json.loads(json_match.group())
except json.JSONDecodeError:
parsed = {"result": raw_output, "format": "text"}
else:
parsed = {"result": raw_output, "format": "text"}
yield {
"type": "final",
"data": parsed
}
except Exception as e:
yield {
"type": "final",
"data": {"error": str(e), "raw_result": str(result_data)}
}
except queue.Empty:
yield {
"type": "error",
"error": "No result received from crew execution"
}