test_live_call.py

"""
test_live_call.py
------------------
Minimal live test: one LLM call, full diagnostics.

Prints:
  - Exact prompt sent
  - Exact JSON body sent to Ollama (intercepted)
  - response.thinking  (empty = thinking mode OFF = fix worked)
  - response.response  (the actual text)
  - Whether JSON parsed correctly

Run with Ollama running:
    python test_live_call.py
"""
import sys, json, time, httpx
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parent))

# ── Intercept HTTP to capture exact request body ──────────────────────────────
original_send = httpx.Client.send
captured_body = {}

def patched_send(self, request, *a, **kw):
    if "/api/generate" in str(request.url):
        try:
            captured_body.update(json.loads(request.content))
        except Exception:
            pass
    return original_send(self, request, *a, **kw)

httpx.Client.send = patched_send

# ── Test prompt (same structure as all our agents) ────────────────────────────
PROMPT = (
    "You are a creative assistant.\n\n"
    "Respond with ONLY this JSON -- no prose, no markdown:\n"
    '{\n  "title": "A short film title",\n  "genre": "one word genre"\n}'
)

print("=" * 60)
print("Scriptify AI -- Live LLM Diagnostic  (generate/raw)")
print("=" * 60)
print(f"\nPrompt ({len(PROMPT)} chars):")
print("-" * 40)
print(PROMPT)
print("-" * 40)

# ── Make the call ─────────────────────────────────────────────────────────────
from backend.config import OLLAMA_MODEL, LLM_TIMEOUT_SECS, LLM_TEMPERATURE, LLM_TOP_P
from backend.agents._llm import _get_client, _extract_json

tokens  = 120
num_ctx = tokens + 600
options = {"temperature": LLM_TEMPERATURE, "top_p": LLM_TOP_P,
           "num_predict": tokens, "num_ctx": num_ctx}

print(f"\nModel    : {OLLAMA_MODEL}")
print(f"Endpoint : generate (raw=True)  -- bypasses modelfile template")
print(f"Options  : {options}")
print(f"\nSending...", flush=True)

t0 = time.time()
response = None
error    = None

try:
    client = _get_client()
    response = client.generate(
        model=OLLAMA_MODEL,
        prompt=PROMPT,
        raw=True,
        options=options,
        keep_alive="10m",
    )
except Exception as e:
    error = e

elapsed = time.time() - t0
print(f"Elapsed  : {elapsed:.2f}s")

# ── Show captured request body ────────────────────────────────────────────────
if captured_body:
    print("\n--- Exact JSON body sent to Ollama ---")
    display = {k: v for k, v in captured_body.items() if k != "prompt"}
    display["prompt_len"] = len(captured_body.get("prompt", ""))
    print(json.dumps(display, indent=2))

if error:
    print(f"\n[ERROR] {type(error).__name__}: {error}")
    if elapsed >= LLM_TIMEOUT_SECS - 2:
        print("\n  TIMEOUT -- model still generating after", LLM_TIMEOUT_SECS, "s")
        print("  Check: ollama ps")
    sys.exit(1)

# ── Inspect response ──────────────────────────────────────────────────────────
thinking_content = getattr(response, "thinking", None) or ""
text_content     = response.response or ""

print("\n--- Response fields ---")
print(f"  .thinking : {repr(thinking_content[:200] if thinking_content else None)}")
print(f"  .response : {repr(text_content[:300])}")

print()
if thinking_content:
    print(f"[FAIL] thinking is NOT empty ({len(thinking_content)} chars)")
    print("       Thinking mode is still ON -- raw=True did not suppress it.")
    print("       Consider using a different model tag or upgrading Ollama.")
else:
    print(f"[OK] thinking is empty/None -- thinking mode OFF")
    print(f"     Response came back in {elapsed:.1f}s")

# ── JSON parse check ──────────────────────────────────────────────────────────
if text_content:
    try:
        parsed = _extract_json(text_content)
        print(f"\n[OK] JSON parsed: {parsed}")
    except Exception as e:
        print(f"\n[WARN] JSON parse failed: {e}")
        print(f"       Raw output: {text_content[:300]}")

print()