nniback / core /categorization /service.py
Kyrlos11's picture
Upload 842 files
22087d0 verified
Raw History Blame Contribute Delete
3.79 kB
"""Service for categorizing projects based on conversation content."""
import json
from typing import List
from core.services.llm import make_llm_api_call
from core.utils.logger import logger
from core.utils.project_helpers import PROJECT_CATEGORIES
MODEL_NAME = "openai/gpt-5-nano-2025-08-07"
async def categorize_from_messages(messages: List[dict]) -> List[str]:
"""
Analyze messages and return applicable categories.
Args:
messages: List of message dicts with 'type' and 'content' (from DB)
Returns:
List of category strings
"""
user_content = _extract_user_content(messages)
if not user_content:
return []
try:
return await _call_llm(user_content)
except Exception as e:
logger.error(f"Categorization LLM call failed: {e}")
return []
def _extract_user_content(messages: List[dict]) -> str:
"""Extract user text messages only.
Message structure:
- type: 'user' (column) with content: {"role": "user", "content": "text"} -> TEXT
- type: 'image_context' with content.content as array -> SKIP
- If content.content is not a string -> SKIP (images, arrays, etc.)
- Skip messages > 5000 chars (likely pasted conversation history)
"""
MAX_MSG_LENGTH = 5000 # Normal user messages are <1000 chars
parts = []
for msg in messages:
# Only 'user' type messages (not image_context, assistant, tool, etc.)
if msg.get('type') != 'user':
continue
content = msg.get('content')
if not content or not isinstance(content, dict):
continue
# Get the inner content - must be a string (not array for images)
inner = content.get('content')
if not isinstance(inner, str):
continue # Skip arrays (images) or other non-string content
text = inner.strip()
# Skip abnormally long messages (likely pasted conversation history)
if len(text) > MAX_MSG_LENGTH:
continue
if text:
parts.append(text)
return "\n".join(parts)
async def _call_llm(content: str) -> List[str]:
"""Call LLM to categorize content."""
system_prompt = f"""You categorize conversations. Return a JSON object with a "categories" key containing ALL applicable categories.
Categories: {', '.join(PROJECT_CATEGORIES)}
Example: {{"categories": ["Research & Information Gathering", "Presentations"]}}
Return ONLY the JSON object."""
response = await make_llm_api_call(
messages=[
{"role": "system", "content": system_prompt},
{"role": "user", "content": f"Categorize:\n\n{content}"}
],
model_name=MODEL_NAME,
max_tokens=1000, # Reasoning models need tokens for chain-of-thought before output
temperature=0.3,
response_format={"type": "json_object"},
stream=False
)
if not response or not response.get('choices'):
return []
raw = response['choices'][0].get('message', {}).get('content', '').strip()
try:
parsed = json.loads(raw)
if isinstance(parsed, list):
categories = parsed
elif isinstance(parsed, dict):
categories = parsed.get('categories') or parsed.get('result') or []
else:
categories = []
# Ensure categories is a list
if not isinstance(categories, list):
categories = []
valid = [c for c in categories if c in PROJECT_CATEGORIES]
return valid if valid else ["Other"]
except json.JSONDecodeError:
logger.warning(f"Failed to parse categorization response: {raw}")
return ["Other"]