test_raw_generate_timing.py

"""
test_raw_generate_timing.py
─────────────────────────────
Isolated test: calls Ollama exactly the way _llm.py does (generate with
raw=True), completely outside FastAPI/Streamlit/LangGraph, and times it.

Run with:
    python test_raw_generate_timing.py

Compare the printed time against your manual `ollama run granite3.3 "..."`
test (~60s). If this script takes roughly the same time, the problem is
elsewhere (timeout config, concurrent load, etc). If this takes much
longer, the raw=True Python path itself is the culprit.
"""

import time
import ollama

from backend.config import OLLAMA_MODEL, OLLAMA_BASE_URL

PROMPT = (
    "Write a JSON object with fields title, genre, synopsis, theme, mood "
    "for a movie idea about a young woman who inherits a detective agency. "
    "Only output JSON, about 150 words total."
)

print(f"Connecting to Ollama at {OLLAMA_BASE_URL}...")
# NOTE: no client-level timeout set here on purpose, so we can see the TRUE
# time the call takes without being cut off early -- this isolates timing
# from timeout configuration entirely.
client = ollama.Client(host=OLLAMA_BASE_URL)

options = {
    "temperature": 0.7,
    "top_p": 0.9,
    "num_predict": 600,
    "num_ctx": 1200,   # same value your idea_agent call used when it failed
}

print(f"Sending raw=True generate() call  [num_predict={options['num_predict']}, num_ctx={options['num_ctx']}]...")
print("(No timeout set -- this will wait as long as it takes. Watch the clock.)")

t0 = time.time()
response = client.generate(
    model=OLLAMA_MODEL,
    prompt=PROMPT,
    raw=True,
    options=options,
    keep_alive="10m",
)
elapsed = time.time() - t0

print(f"\n{'=' * 60}")
print(f"DONE in {elapsed:.1f} seconds")
print(f"{'=' * 60}")
print(f"\nResponse text ({len(response.response)} chars):")
print(response.response)

thinking = getattr(response, "thinking", None)
if thinking:
    print(f"\n[WARNING] Thinking content present ({len(thinking)} chars) -- not suppressed!")