Add Ollama backend + Qwen3 local inference support

- Extractor now supports two backends: ollama (local) and anthropic (cloud)
- Default is ollama with qwen3:14b (fits 16GB VRAM)
- Set num_ctx to 32768 for full-script processing
- Added --backend and --ollama-url CLI flags
- Added The Last Backup test script
- Tested: 12/12 scenes valid on dialogue_heavy, 12/13 on the_last_backup

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

requirements.txt

@@ -1,3 +1,4 @@
 anthropic>=0.40.0
 pydantic>=2.0.0
 python-dotenv>=1.0.0
+requests>=2.31.0

run.py

@@ -14,7 +14,9 @@ def main():
     parser = argparse.ArgumentParser(description="AI Movie Production Pipeline — Phase 1")
     parser.add_argument("--script", type=str, help="Path to .fountain script file")
     parser.add_argument("--project", type=str, help="Project name (determines output directory)")
-    parser.add_argument("--model", type=str, default="claude-sonnet-4-20250514", help="Model ID")
+    parser.add_argument("--model", type=str, default="qwen3:14b", help="Model ID (default: qwen3:14b)")
+    parser.add_argument("--backend", type=str, default="ollama", choices=["ollama", "anthropic"], help="AI backend (default: ollama)")
+    parser.add_argument("--ollama-url", type=str, default="http://localhost:11434", help="Ollama server URL")
     parser.add_argument("--scene", type=int, default=None, help="Process only this scene number")
     parser.add_argument("--dry-run", action="store_true", help="Validate inputs only, no AI calls")
     parser.add_argument("--force", action="store_true", help="Ignore cache, re-run even if unchanged")
@@ -24,14 +26,14 @@ def main():
     args = parser.parse_args()
 
     if args.test:
-        run_tests(args.model, args.output_dir)
+        run_tests(args.model, args.backend, args.ollama_url, args.output_dir)
         return
 
     if not args.script or not args.project:
         parser.error("--script and --project are required (unless using --test)")
 
-    api_key = os.environ.get("ANTHROPIC_API_KEY")
+    api_key = os.environ.get("ANTHROPIC_API_KEY", "")
-    if not api_key and not args.dry_run:
+    if args.backend == "anthropic" and not api_key and not args.dry_run:
         print("ERROR: ANTHROPIC_API_KEY not set. Set it in .env or environment.")
         sys.exit(1)
 
@@ -44,8 +46,10 @@ def main():
     result = run_phase1(
         script_path=args.script,
         project_name=args.project,
-        api_key=api_key or "",
+        api_key=api_key,
         model=args.model,
+        backend=args.backend,
+        ollama_url=args.ollama_url,
         output_dir=args.output_dir,
         scene_filter=args.scene,
         dry_run=args.dry_run,
@@ -60,11 +64,11 @@ def main():
     sys.exit(0)
 
 
-def run_tests(model: str, output_dir: str):
+def run_tests(model: str, backend: str, ollama_url: str, output_dir: str):
     """Run test suite against all scripts in test_scripts/."""
-    api_key = os.environ.get("ANTHROPIC_API_KEY")
+    api_key = os.environ.get("ANTHROPIC_API_KEY", "")
-    if not api_key:
+    if backend == "anthropic" and not api_key:
-        print("ERROR: ANTHROPIC_API_KEY required for tests")
+        print("ERROR: ANTHROPIC_API_KEY required for anthropic backend tests")
         sys.exit(1)
 
     test_dir = os.path.join(os.path.dirname(__file__), "test_scripts")
@@ -93,6 +97,8 @@ def run_tests(model: str, output_dir: str):
             project_name=project_name,
             api_key=api_key,
             model=model,
+            backend=backend,
+            ollama_url=ollama_url,
             output_dir=output_dir,
         )
 

src/execution/runner.py

@@ -34,8 +34,10 @@ class PipelineResult:
 def run_phase1(
     script_path: str,
     project_name: str,
-    api_key: str,
+    api_key: str = "",
-    model: str = "claude-sonnet-4-20250514",
+    model: str = "qwen3:14b",
+    backend: str = "ollama",
+    ollama_url: str = "http://localhost:11434",
     output_dir: str = "output",
     scene_filter: int | None = None,
     dry_run: bool = False,
@@ -46,8 +48,10 @@ def run_phase1(
     Args:
         script_path: Path to .fountain file.
         project_name: Project name for output directory.
-        api_key: Anthropic API key.
+        api_key: API key (required for anthropic backend).
         model: Model ID.
+        backend: "ollama" or "anthropic".
+        ollama_url: Ollama server URL.
         output_dir: Base output directory.
         scene_filter: If set, only process this scene number in Layer 2.
         dry_run: If True, validate inputs only — no AI calls.
@@ -108,7 +112,7 @@ def run_phase1(
     for chunk_idx, chunk in enumerate(chunks):
         # Extract scenes from this chunk
         def do_extract(data):
-            return extract_scenes(data, contract_path, api_key, model)
+            return extract_scenes(data, contract_path, api_key, model, backend, ollama_url)
 
         run_id = logger.start("L2", scene_id=chunk_idx)
         chunk_input_hash = f"sha256:{hashlib.sha256(json.dumps([e.model_dump() for e in chunk.elements]).encode()).hexdigest()}"

src/understanding/extractor.py

@@ -1,8 +1,15 @@
-"""Layer 2 AI extractor — sends normalized script to Claude, receives scene JSON."""
+"""Layer 2 AI extractor — sends normalized script to LLM, receives scene JSON.
+
+Supports two backends:
+- Ollama (local, via OpenAI-compatible API)
+- Anthropic (cloud, via Anthropic SDK)
+
+Backend is selected by the `backend` parameter or auto-detected from model name.
+"""
 
 import json
+import requests
 from dataclasses import dataclass
-from anthropic import Anthropic
 from src.schemas.normalized_script import NormalizedScript
 
 
@@ -19,16 +26,20 @@ class ExtractionError(Exception):
 def extract_scenes(
     script: NormalizedScript,
     contract_path: str,
-    api_key: str,
+    api_key: str = "",
-    model: str = "claude-sonnet-4-20250514",
+    model: str = "qwen3:14b",
+    backend: str = "ollama",
+    ollama_url: str = "http://localhost:11434",
 ) -> ExtractionResult:
     """Extract structured scene data from a normalized script using AI.
 
     Args:
         script: Validated NormalizedScript from Layer 1.
         contract_path: Path to the prompt contract JSON file.
-        api_key: Anthropic API key.
+        api_key: API key (required for anthropic backend, ignored for ollama).
-        model: Model ID to use.
+        model: Model ID.
+        backend: "ollama" or "anthropic".
+        ollama_url: Ollama server URL (default localhost:11434).
 
     Returns:
         ExtractionResult with raw scene dicts and token usage.
@@ -46,32 +57,31 @@ def extract_scenes(
     # Render user prompt
     user_prompt = contract["user_prompt_template"].replace("{{script_text}}", script_text)
 
-    # Call Claude API
+    if backend == "ollama":
-    client = Anthropic(api_key=api_key)
+        response_text, token_usage = _call_ollama(
-    response = client.messages.create(
+            model, contract["system_prompt"], user_prompt,
-        model=model,
+            contract["max_output_tokens"], ollama_url,
-        max_tokens=contract["max_output_tokens"],
-        temperature=0,
-        system=contract["system_prompt"],
-        messages=[{"role": "user", "content": user_prompt}],
         )
-
+    elif backend == "anthropic":
-    # Extract text content
+        response_text, token_usage = _call_anthropic(
-    response_text = response.content[0].text
+            model, contract["system_prompt"], user_prompt,
+            contract["max_output_tokens"], api_key,
+        )
+    else:
+        raise ExtractionError(f"Unknown backend: {backend}")
 
     # Parse JSON
     try:
         parsed = json.loads(response_text)
     except json.JSONDecodeError as e:
-        # Try to extract JSON from response if wrapped in markdown
         cleaned = _extract_json(response_text)
         if cleaned:
             try:
                 parsed = json.loads(cleaned)
             except json.JSONDecodeError:
-                raise ExtractionError(f"AI response is not valid JSON: {e}") from e
+                raise ExtractionError(f"AI response is not valid JSON: {e}\nResponse:\n{response_text[:500]}") from e
         else:
-            raise ExtractionError(f"AI response is not valid JSON: {e}") from e
+            raise ExtractionError(f"AI response is not valid JSON: {e}\nResponse:\n{response_text[:500]}") from e
 
     # Extract scenes array
     if isinstance(parsed, dict) and "scenes" in parsed:
@@ -84,12 +94,73 @@ def extract_scenes(
     if not isinstance(scenes, list):
         raise ExtractionError(f"'scenes' is not a list: {type(scenes)}")
 
+    return ExtractionResult(raw_scenes=scenes, token_usage=token_usage)
+
+
+def _call_ollama(
+    model: str, system_prompt: str, user_prompt: str,
+    max_tokens: int, ollama_url: str,
+) -> tuple[str, dict]:
+    """Call Ollama's chat API."""
+    payload = {
+        "model": model,
+        "messages": [
+            {"role": "system", "content": system_prompt},
+            {"role": "user", "content": user_prompt},
+        ],
+        "stream": False,
+        "options": {
+            "temperature": 0,
+            "num_predict": max_tokens,
+            "num_ctx": 32768,
+        },
+        "format": "json",
+    }
+
+    try:
+        resp = requests.post(f"{ollama_url}/api/chat", json=payload, timeout=600)
+        resp.raise_for_status()
+    except requests.RequestException as e:
+        raise ExtractionError(f"Ollama API call failed: {e}") from e
+
+    data = resp.json()
+    response_text = data.get("message", {}).get("content", "")
+
+    # Ollama provides token counts
+    token_usage = {
+        "input": data.get("prompt_eval_count", 0),
+        "output": data.get("eval_count", 0),
+    }
+
+    if not response_text:
+        raise ExtractionError("Ollama returned empty response")
+
+    return response_text, token_usage
+
+
+def _call_anthropic(
+    model: str, system_prompt: str, user_prompt: str,
+    max_tokens: int, api_key: str,
+) -> tuple[str, dict]:
+    """Call Anthropic's messages API."""
+    from anthropic import Anthropic
+
+    client = Anthropic(api_key=api_key)
+    response = client.messages.create(
+        model=model,
+        max_tokens=max_tokens,
+        temperature=0,
+        system=system_prompt,
+        messages=[{"role": "user", "content": user_prompt}],
+    )
+
+    response_text = response.content[0].text
     token_usage = {
         "input": response.usage.input_tokens,
         "output": response.usage.output_tokens,
     }
 
-    return ExtractionResult(raw_scenes=scenes, token_usage=token_usage)
+    return response_text, token_usage
 
 
 def _elements_to_text(script: NormalizedScript) -> str:

test_scripts/the_last_backup.fountain

@@ -0,0 +1,431 @@
+Title: The Last Backup
+Credit: Written for pipeline testing
+Author: OpenAI
+Draft date: 2026-04-06
+
+INT. SERVER ROOM - NIGHT
+
+Rows of aging servers hum in blue light. A red STATUS LED blinks on the far rack.
+
+MARA REYES, 38, sharp, exhausted, still in work clothes, stands over an open terminal. She grips a paper notebook filled with handwritten commands.
+
+On the screen: BACKUP FAILED.
+
+MARA
+No. No, no, no.
+
+She types fast, reruns the job, watches.
+
+The screen flashes the same error.
+
+A metal KEY on a red tag sits beside the keyboard. A half-empty coffee cup trembles from the vibration of the cooling fans.
+
+MARA
+Come on. Just hold together one more night.
+
+Her phone buzzes. Caller ID: ELI.
+
+She ignores it.
+
+INT. APARTMENT KITCHEN - NIGHT
+
+A small apartment. Functional, cluttered, dim.
+
+ELI REYES, 16, hoodie, anxious energy, stands at the counter eating cereal from the box. His phone is on speaker.
+
+ELI
+Mom?
+
+INTERCUT PHONE CALL
+
+MARA paces in the server room, still staring at the terminal.
+
+MARA
+I'm here.
+
+ELI
+You said you'd be home before ten.
+
+MARA
+I know.
+
+ELI
+You always say that when something's on fire.
+
+MARA
+Nothing's on fire.
+
+A beat. A LOUD POP from the server room.
+
+MARA closes her eyes.
+
+MARA
+Something may be overheating.
+
+ELI
+That's worse.
+
+INT. SERVER ROOM - LATER
+
+The STATUS LED is now solid red.
+
+MARA has removed her blazer. Sleeves rolled. Coffee gone cold.
+
+She flips open the notebook. A page is marked: "MANUAL FAILOVER - LAST RESORT."
+
+At the rack, she inserts the red-tagged KEY into a locked panel.
+
+Inside: one small drive bay. Empty.
+
+MARA
+You have got to be kidding me.
+
+Her phone buzzes again. This time: VOICEMAIL RECEIVED.
+
+She doesn't listen.
+
+EXT. PARKING LOT - NIGHT
+
+Rain slicks the pavement outside the low industrial building.
+
+JONAH VALE, 40s, maintenance contractor, steps out of an old pickup truck carrying a dented toolbox and a flashlight.
+
+He looks up at the building like it insulted his family.
+
+JONAH
+Every emergency job is just somebody else's procrastination with weather.
+
+He heads inside.
+
+INT. SERVER ROOM - NIGHT
+
+JONAH kneels by an exposed side panel while MARA hovers nearby.
+
+JONAH
+This system should've been retired five years ago.
+
+MARA
+It was supposed to be replaced last quarter.
+
+JONAH
+And yet here we are, praying to obsolete metal.
+
+He shines the flashlight into the rack.
+
+JONAH (CONT'D)
+You got a spare drive?
+
+MARA says nothing.
+
+JONAH looks at the empty bay. Then at her.
+
+JONAH (CONT'D)
+That's a no.
+
+MARA
+I had one.
+
+JONAH
+Past tense is doing a lot of work there.
+
+He stands, wipes his hands.
+
+JONAH (CONT'D)
+Best case, we stabilize it long enough to image what's left.
+
+MARA
+Worst case?
+
+JONAH
+You already know the worst case or you wouldn't be shaking.
+
+MARA looks down. She is shaking.
+
+INT. APARTMENT KITCHEN - SAME NIGHT
+
+ELI sits at the table now. The cereal is soggy. The overhead light flickers.
+
+He opens his laptop. On screen: a school form titled EMERGENCY CONTACT UPDATE.
+
+The cursor blinks next to "SECONDARY CONTACT."
+
+He types: JONAH VALE
+Then deletes it.
+
+He types: NONE
+Then stops.
+
+ELI
+(to himself)
+Great.
+
+He closes the laptop.
+
+INT. SERVER ROOM - LATER
+
+A portable WORK LIGHT now casts harsh white shadows.
+
+JONAH has rigged a temporary cooling fan with zip ties. MARA sits at the terminal.
+
+The system begins a recovery scan.
+
+ON SCREEN: 12%... 13%... 14%
+
+MARA almost doesn't breathe.
+
+JONAH
+Don't look at it like that.
+
+MARA
+Like what?
+
+JONAH
+Like fear improves machinery.
+
+MARA
+It improves people.
+
+JONAH
+Debatable.
+
+Her phone buzzes again. She finally listens to the voicemail.
+
+ELI (V.O.)
+Hey. Sorry. I know you're working. Just... if you can call when you get a second, call, okay?
+
+MARA stares at nothing.
+
+MARA
+I missed his concert.
+
+JONAH
+Tonight?
+
+MARA nods.
+
+JONAH
+That explains the face.
+
+MARA
+What face?
+
+JONAH
+The one that says the server isn't the only thing failing.
+
+A long beat.
+
+ON SCREEN: 33%
+
+INT. APARTMENT BEDROOM - NIGHT
+
+ELI lies on his bed fully dressed, headphones on, staring at the ceiling.
+
+A trophy sits on the dresser beside a printed concert program.
+
+He picks up the program, folds it once, then again.
+
+His room is neat in the way control freaks keep it neat when other things aren't.
+
+His phone lights up: MOM CALLING.
+
+He lets it ring.
+
+INT. SERVER ROOM - NIGHT
+
+ON SCREEN: 67%
+
+MARA leaves a voicemail.
+
+MARA
+Eli. I know you're awake. I know you don't want to hear from me right now, but I need you to hear this part. I am sorry. Not work-sorry. Not almost-sorry. Just sorry.
+
+JONAH pretends not to listen.
+
+MARA (CONT'D)
+When this is done, I'm coming home.
+
+She hangs up.
+
+JONAH
+Good voicemail.
+
+MARA
+You grading me now?
+
+JONAH
+No. If I was grading you, I'd mention you still haven't asked the important question.
+
+MARA
+Which is?
+
+JONAH
+If the recovered data has somewhere to go.
+
+Silence.
+
+MARA turns slowly toward the empty drive bay.
+
+MARA
+Right.
+
+JONAH
+Please tell me you have external storage.
+
+MARA reaches into her bag and pulls out a rugged BLACK PORTABLE SSD.
+
+JONAH exhales.
+
+JONAH
+See? Miracles. Tiny, overpriced miracles.
+
+INT. SERVER ROOM - LATER
+
+The BLACK PORTABLE SSD is connected.
+
+ON SCREEN: RECOVERY COMPLETE - PARTIAL IMAGE AVAILABLE.
+
+MARA closes her eyes in relief.
+
+Then another message appears.
+
+CORRUPTION DETECTED IN ARCHIVE SEGMENT 03.
+
+MARA
+Of course.
+
+JONAH
+How bad?
+
+MARA scans the report.
+
+MARA
+Not core infrastructure. Historical footage. Client archive.
+
+JONAH
+Can it be rebuilt?
+
+MARA
+Some of it.
+
+JONAH
+Then tonight is not the apocalypse. Just a smaller religion.
+
+MARA laughs despite herself. First time all night.
+
+INT. APARTMENT KITCHEN - PRE-DAWN
+
+Gray light leaks through the blinds.
+
+ELI sits at the table again, half asleep. The folded concert program is beside him.
+
+The front door unlocks.
+
+MARA steps inside carrying her bag, soaked from the rain, wrecked but upright.
+
+A long silence.
+
+MARA
+I saved most of it.
+
+ELI
+Congratulations.
+
+MARA sets the bag down.
+
+MARA
+I know that's not the point.
+
+ELI
+No, it kind of never is.
+
+She nods. Takes the hit.
+
+MARA
+I brought breakfast.
+
+She lifts a paper bag. It's crushed and slightly wet.
+
+ELI almost smiles.
+
+ELI
+That bag has been through combat.
+
+MARA
+So have I.
+
+Another silence. Softer now.
+
+ELI notices the BLACK PORTABLE SSD sticking out of her bag.
+
+ELI
+Is that the thing that kept you there?
+
+MARA
+Part of it.
+
+ELI
+You always say it's temporary.
+
+MARA
+I know.
+
+ELI
+And is it?
+
+MARA looks at him. This time she answers carefully.
+
+MARA
+No. Not if I keep pretending it fixes itself.
+
+He studies her, deciding whether to believe it.
+
+ELI
+Okay.
+
+Not forgiveness. Not yet. But not nothing.
+
+INT. SERVER ROOM - MORNING
+
+Golden morning light cuts through a high window.
+
+The red STATUS LED is off. The temporary cooling rig still hangs in place, ugly and effective.
+
+A printed sign is taped to the rack:
+
+OUT OF SERVICE - DO NOT RELY ON LEGACY NODE
+
+Below it, the red-tagged KEY hangs from a hook.
+
+The BLACK PORTABLE SSD is gone.
+
+JONAH enters alone with coffee. He reads the sign.
+
+JONAH
+Look at that. Growth.
+
+He places a new replacement drive on the shelf below the empty bay.
+
+JONAH (CONT'D)
+Now you show up.
+
+He exits.
+
+EXT. BUILDING ROOF - MORNING
+
+MARA stands alone on the roof with her phone, city waking up around her.
+
+She opens a new note and types:
+
+1. Replace legacy node
+2. Build real backup path
+3. Go to Eli's next concert
+
+She looks at the list.
+
+Then adds:
+
+4. Leave work before it becomes an emergency
+
+She saves it.
+
+FADE OUT.