Remove OpenAI Cleanup Stage From The Pipeline

README.md

@@ -9,7 +9,7 @@ python_version: "3.12"
 app_file: app.py
 pinned: false
 license: mit
-short_description: End-to-end Parakeet + Pyannote + OpenAI transcript pipeline
+short_description: End-to-end Parakeet + Pyannote transcript pipeline
 ---
 
 This Space is optimized for API usage on ZeroGPU.
@@ -19,7 +19,6 @@ Single production pipeline:
 1. Parakeet transcription (word timestamps) on ZeroGPU
 2. Pyannote Community-1 diarization on ZeroGPU
 3. Merge transcript + diarization (CPU)
-4. OpenAI speaker inference + cleanup (CPU)
 
 Model setup is global/outside `@spaces.GPU` so setup time is not billed to ZeroGPU windows.
 
@@ -30,10 +29,10 @@ Model setup is global/outside `@spaces.GPU` so setup time is not billed to ZeroG
 ## `/run_complete_pipeline` inputs
 - `audio_file` (file path from Gradio client upload)
 - `huggingface_token`
-- `openai_api_key`
-- `executive_names_csv`
+- `openai_api_key` (accepted for compatibility, unused in Space)
+- `executive_names_csv` (accepted for compatibility, unused in Space)
 
-Returns: completed transcript JSON only.
+Returns: merged transcript JSON only.
 
 ## `/get_debug_output` inputs
 - `run_id` (optional)
@@ -42,7 +41,6 @@ Returns: raw/debug payload for the latest run (or specific run if provided), inc
 - Parakeet raw output + ZeroGPU timing
 - Pyannote raw output + ZeroGPU timing
 - merged transcript payload
-- OpenAI raw responses + token usage
 - aggregated timing
 
 ## IPython example
@@ -54,12 +52,12 @@ AUDIO_FILE = "Q3-FY26_5min.mp3"
 
 client = Client(SPACE)
 
-# Run end-to-end pipeline
-final_transcript = client.predict(
+# Run end-to-end pipeline (returns merged transcript)
+merged_transcript = client.predict(
     audio_file=handle_file(AUDIO_FILE),
     huggingface_token="hf_xxx",
-    openai_api_key="sk-xxx",
-    executive_names_csv="Ashok Vaswani,Devang Gheewala,Kotak 811,CASA,NIM",
+    openai_api_key="",       # unused
+    executive_names_csv="",  # unused
     api_name="/run_complete_pipeline",
 )
 

app.py

@@ -12,7 +12,6 @@ from src.diarization_service import run_chunked_diarization
 from src.merge_service import merge_parakeet_pyannote_outputs
 from src.models.parakeet_model import preload_parakeet_model, run_parakeet
 from src.models.pyannote_community_model import preload_pyannote_pipeline, run_pyannote_community_chunk
-from src.openai_cleanup_service import run_openai_cleanup_pipeline
 from src.utils import get_audio_duration_seconds
 
 # Suppress a known deprecation warning emitted by a transitive dependency in spaces.
@@ -58,14 +57,11 @@ def _store_debug_payload(payload: dict[str, Any]) -> str:
 def _parse_main_request(
     audio_file: str | None,
     huggingface_token: str | None,
-    openai_api_key: str | None,
 ) -> None:
     if audio_file is None:
         raise gr.Error("No audio file submitted. Upload an audio file first.")
     if not huggingface_token or not huggingface_token.strip():
         raise gr.Error("huggingface_token is required for pyannote/speaker-diarization-community-1.")
-    if not openai_api_key or not openai_api_key.strip():
-        raise gr.Error("openai_api_key is required for speaker inference and cleanup.")
 
 
 # Global setup (outside @spaces.GPU) so setup cost is not charged to ZeroGPU inference window.
@@ -116,7 +112,10 @@ def run_complete_pipeline(
     openai_api_key: str,
     executive_names_csv: str,
 ):
-    _parse_main_request(audio_file, huggingface_token, openai_api_key)
+    # Kept in signature for compatibility with existing clients; not used on Space.
+    _ = openai_api_key
+    _ = executive_names_csv
+    _parse_main_request(audio_file, huggingface_token)
     _raise_preload_error_if_any(PARAKEET_V3)
 
     started_at = time.perf_counter()
@@ -159,18 +158,6 @@ def run_complete_pipeline(
         diarization_key="exclusive_speaker_diarization",
     )
 
-    # 4) OpenAI speaker inference + cleanup outside ZeroGPU.
-    openai_result = run_openai_cleanup_pipeline(
-        merged_transcript=merged_transcript,
-        openai_api_key=openai_api_key,
-        executive_names_csv=executive_names_csv,
-        cleanup_model="gpt-5",
-        timeout_seconds=600.0,
-        max_turns_per_chunk=80,
-        max_chars_per_chunk=22000,
-    )
-    cleaned_transcript = openai_result["cleaned_transcript"]
-
     total_gpu_window_seconds = float(parakeet_response["zerogpu_timing"].get("gpu_window_seconds", 0.0)) + float(
         pyannote_response.get("zerogpu_timing", {}).get("gpu_window_seconds", 0.0)
     )
@@ -187,20 +174,16 @@ def run_complete_pipeline(
         },
         "inputs": {
             "audio_file": str(audio_file),
-            "executive_names_csv": executive_names_csv or "",
             "huggingface_token_provided": bool(huggingface_token),
-            "openai_api_key_provided": bool(openai_api_key),
         },
         "parakeet_response": parakeet_response,
         "pyannote_response": pyannote_response,
         "merged_transcript": merged_transcript,
-        "openai_debug": openai_result.get("debug", {}),
-        "final_transcript": cleaned_transcript,
     }
     _store_debug_payload(debug_payload)
 
-    # Return only final transcript JSON for production API.
-    return cleaned_transcript
+    # Return merged transcript JSON (OpenAI cleanup is intentionally local/off-space).
+    return merged_transcript
 
 
 def get_debug_output(run_id: str | None):
@@ -215,10 +198,10 @@ def get_debug_output(run_id: str | None):
     return {"run_id": _LAST_DEBUG_RUN_ID, "debug": _DEBUG_RUNS[_LAST_DEBUG_RUN_ID]}
 
 
-with gr.Blocks(title="Parakeet + Pyannote + OpenAI Pipeline") as demo:
+with gr.Blocks(title="Parakeet + Pyannote Pipeline") as demo:
     gr.Markdown(
         "# End-to-end transcript pipeline\n"
-        "Runs Parakeet transcription, Pyannote diarization, merge, then OpenAI speaker mapping + cleanup."
+        "Runs Parakeet transcription, Pyannote diarization, then merges into a combined transcript JSON."
     )
 
     audio_file = gr.Audio(
@@ -230,17 +213,11 @@ with gr.Blocks(title="Parakeet + Pyannote + OpenAI Pipeline") as demo:
         label="HuggingFace token",
         type="password",
     )
-    openai_api_key = gr.Textbox(
-        label="OpenAI API key",
-        type="password",
-    )
-    executive_names_csv = gr.Textbox(
-        label="Executive names / terms (comma-separated)",
-        placeholder="Ashok Vaswani,Devang Gheewala,Kotak 811,CASA,NIM",
-    )
+    openai_api_key = gr.Textbox(label="OpenAI API key (unused in Space)", type="password")
+    executive_names_csv = gr.Textbox(label="Executive names / terms (unused in Space)")
 
     run_btn = gr.Button("Run full pipeline")
-    output = gr.JSON(label="Completed transcript JSON")
+    output = gr.JSON(label="Combined transcript JSON")
 
     run_btn.click(
         fn=run_complete_pipeline,

requirements.txt

@@ -8,4 +8,3 @@ transformers
 accelerate
 nemo_toolkit[asr]
 pyannote.audio
-openai