test_models.py - main - Paff/declarant

test_models.py
63 строки3 KB
1pash1985-gif
Unified doc list UI + AI prompt for FORM_A
05 авг 2026, 10:09
05 авг 2026, 10:09e439f08
"""Test different models and approaches for the reasoning token issue."""
import sys, os, re, json, time
sys.path.insert(0, '/app/web')
os.chdir('/app/web')

from ai_parser import extract_text_from_excel, client, INVOICE_SYSTEM_PROMPT, _get_corrections_context
from config import DEEPSEEK_API_KEY, DEEPSEEK_BASE_URL
from openai import OpenAI

xls_path = '/app/web/uploads/reports/2026-07-27/444/sources/invoice_444_1.xls'
raw_text = extract_text_from_excel(xls_path)

# Strip URLs
url_pattern = r'https?://\S+'
text_clean = re.sub(url_pattern, '', raw_text)
text_clean = re.sub(r'  +', ' ', text_clean)
text_clean = re.sub(r' \| $', '', text_clean)

print(f'Raw text: {len(raw_text)} chars')
print(f'Clean text: {len(text_clean)} chars')
print('---')

system_prompt = INVOICE_SYSTEM_PROMPT + _get_corrections_context()

# Models to test
models = ['deepseek-chat', 'deepseek-v4-pro', 'deepseek-v4-flash']
texts = [('clean', text_clean)]

for model_name in models:
    for text_label, text in texts:
        print(f'=== {model_name} + {text_label} text (max_tokens=8192) ===')
        t0 = time.time()
        try:
            response = client.chat.completions.create(
                model=model_name,
                messages=[
                    {'role': 'system', 'content': system_prompt},
                    {'role': 'user', 'content': f'Вот текст инвойса:\n\n{text}'},
                ],
                temperature=0.1,
                max_tokens=8192,
                response_format={'type': 'json_object'},
            )
            elapsed = time.time() - t0
            fr = response.choices[0].finish_reason
            content = response.choices[0].message.content or ''
            cl = len(content)
            ct = response.usage.completion_tokens if response.usage else 0
            rt = '?'
            if response.usage and hasattr(response.usage, 'completion_tokens_details'):
                rt = getattr(response.usage.completion_tokens_details, 'reasoning_tokens', 0) or 0
            print(f'  finish_reason={fr}, content_len={cl}, completion={ct}, reasoning={rt}, time={elapsed:.1f}s')
            if cl > 0:
                print(f'  Preview: {content[:200]}')
                try:
                    parsed = json.loads(content)
                    print(f'  PARSED OK! supplier={parsed.get("supplier_name", "N/A")}')
                except:
                    print(f'  JSON parse failed')
        except Exception as e:
            elapsed = time.time() - t0
            print(f'  ERROR ({elapsed:.1f}s): {e}')
        print()