Dev

.github/workflows/backend.yml

@@ -42,7 +42,8 @@ jobs:
     - uses: actions/checkout@v4
       with:
         token: ${{ secrets.GITHUB_TOKEN }}
-
+        persist-credentials: true
+
     - name: Set up Python
       uses: actions/setup-python@v5
       with:
@@ -69,11 +70,21 @@ jobs:
       run: |
         git diff --exit-code || echo "formatted=true" >> $GITHUB_OUTPUT
 
+    - name: Import GPG key
+      if: steps.verify_diff.outputs.formatted == 'true' && github.event_name == 'push'
+      uses: crazy-max/ghaction-import-gpg@v6
+      with:
+        gpg_private_key: ${{ secrets.GPG_PRIVATE_KEY }}
+        passphrase: ${{ secrets.GPG_PASSPHRASE }}
+        git_user_signingkey: true
+        git_commit_gpgsign: true
+        git_config_global: true
+
     - name: Commit and push if changed
       if: steps.verify_diff.outputs.formatted == 'true' && github.event_name == 'push'
       run: |
         git config --global user.name 'github-actions[bot]'
-        git config --global user.email 'github-actions[bot]@users.noreply.github.com'
+        git config --global user.email 'noreply@github.com'
         git add -A
-        git commit -m "style: auto-format Python code with black and isort"
+        git commit -S -m "style: auto-format Python code with black and isort"
         git push

backend/.env.example

@@ -6,12 +6,23 @@ OLLAMA_SERVER_PORT=YOUR_PORT
 LM_STUDIO_SERVER_URL=YOUR_LM_STUDIO_SERVER_URL_HERE
 LM_STUDIO_SERVER_PORT=YOUR_PORT
 
-# Change to either 'ollama' or 'lm_studio' depending on which LLM server you want to use
+# Change to either 'ollama', 'lm_studio', or 'cloud' depending on which LLM server you want to use
 LLM_PROVIDER=YOUR_LLM_PROVIDER_HERE
 
 # Change to the model name you want to use from your LLM server (ensure the model is already installed on the server)
 LLM_MODEL_NAME=YOUR_LLM_MODEL_NAME_HERE
 
+# Cloud API settings (only needed if using 'cloud' provider for LLM or transcription)
+# Use OpenAI API or any OpenAI-compatible API (e.g., Together AI, OpenRouter, Azure OpenAI, etc.)
+CLOUD_API_BASE_URL=https://api.openai.com/v1
+CLOUD_API_KEY=YOUR_API_KEY_HERE
+
+# Transcription settings
+# Change to either 'local' or 'cloud' - local uses Whisper models, cloud uses OpenAI-compatible API
+TRANSCRIPTION_PROVIDER=local
+# Model name to use for cloud transcription (whisper-large-v3 is default, OpenAI uses whisper-1)
+TRANSCRIPTION_MODEL_NAME=whisper-large-v3
+
 # Authorization header for API access
 # To generate a new secure token, run: openssl rand -hex 32
 API_AUTH_TOKEN=YOUR_AUTH_TOKEN_HERE

backend/pyproject.toml

@@ -9,6 +9,7 @@ dependencies = [
     "faster-whisper>=1.2.1",
     "mlx-whisper>=0.4.3",
     "python-dotenv>=1.2.1",
+    "requests>=2.32.0",
     "torch>=2.8.0",
     "uvicorn>=0.39.0",
 ]

backend/services/text_cleaner.py

@@ -7,22 +7,42 @@
 
 llm_provider = os.getenv("LLM_PROVIDER")
 llm_model_name = os.getenv("LLM_MODEL_NAME")
+cloud_api_base_url = os.getenv("CLOUD_API_BASE_URL")
+cloud_api_key = os.getenv("CLOUD_API_KEY")
 
-# Define the system prompt for disfluency removal, not recommended to modify
-system_prompt = (
-    "You are a text normalization system for spoken transcripts. "
-    "Your task is disfluency removal only. "
-    "Remove filler words (e.g., 'uh', 'um', 'like'), false starts, repetitions, "
-    "and self-corrections while preserving the original meaning, intent, and tone. "
-    "Do not summarize or paraphrase unless required to remove a disfluency. "
-    "Do not add or infer new information. "
-    "Preserve punctuation, capitalization, technical terms, acronyms, "
-    "and domain-specific language."
-    "Output only the cleaned transcript without any additional commentary or explanations."
-)
-
-
-def remove_disfluencies(text):
+# Define the system prompt for text cleaning task
+system_prompt = """You are a transcript cleaning system that removes disfluencies from spoken text.
+
+Your task:
+1. Remove filler words: "uh", "um", "like", "ahh", etc.
+2. Remove false starts and self-corrections - keep ONLY the final intended statement
+3. Remove phrases like "actually", "wait no", "scratch that", "I meant" when they signal corrections
+4. Preserve all technical terms EXACTLY as spoken (npm, PostgreSQL, React, Socket.io, etc.)
+5. Preserve the speaker's final intent and meaning
+
+Examples:
+Input: "I want to use Docker and Kubernetes. Actually no, just Docker."
+Output: "I want to use Docker."
+
+Input: "Let's use MongoDB. Let's use Postgres instead."
+Output: "Let's use Postgres."
+
+Input: "I want WebSockets. Wait no, I meant Socket.io and Redis."
+Output: "I want Socket.io and Redis."
+
+Input: "We need npm packages like React, Next.js, and Tailwind CSS."
+Output: "We need npm packages like React, Next.js, and Tailwind CSS."
+
+Critical rules:
+- When someone changes their mind ("actually no", "instead", "wait no"), ONLY keep the final decision
+- Do NOT keep both the old and new choices
+- Technical terms are sacred - never change them
+- Output ONLY the cleaned text, no explanations
+
+Clean this transcript:"""
+
+
+def clean_text(text):
     """Sends the input text to an external AI model for disfluency removal."""
     if llm_provider == "lm_studio":
         api_url = os.getenv("LM_STUDIO_SERVER_URL")
@@ -32,6 +52,13 @@ def remove_disfluencies(text):
         api_url = os.getenv("OLLAMA_SERVER_URL")
         api_port = os.getenv("OLLAMA_SERVER_PORT")
         endpoint = f"{api_url}:{api_port}/api/chat"
+    elif llm_provider == "cloud":
+        # Ensure base URL ends with /chat/completions for OpenAI-compatible APIs
+        base_url = cloud_api_base_url.rstrip("/")
+        if not base_url.endswith("/chat/completions"):
+            endpoint = f"{base_url}/chat/completions"
+        else:
+            endpoint = base_url
     else:
         raise ValueError(f"Unsupported LLM provider: {llm_provider}")
 
@@ -46,12 +73,19 @@ def remove_disfluencies(text):
             "max_tokens": 2000,
         }
 
-        response = requests.post(endpoint, json=payload, timeout=180)
+        headers = {}
+        if llm_provider == "cloud":
+            if not cloud_api_key:
+                raise ValueError("CLOUD_API_KEY must be set when using cloud provider")
+            headers["Authorization"] = f"Bearer {cloud_api_key}"
+            headers["Content-Type"] = "application/json"
+
+        response = requests.post(endpoint, json=payload, headers=headers, timeout=180)
         response.raise_for_status()
 
         result = response.json()
 
-        if llm_provider == "lm_studio":
+        if llm_provider == "lm_studio" or llm_provider == "cloud":
             cleaned_text = result["choices"][0]["message"]["content"]
         elif llm_provider == "ollama":
             cleaned_text = result["message"]["content"]
@@ -68,3 +102,5 @@ def remove_disfluencies(text):
         raise Exception(f"Error calling LLM API: {str(e)}")
     except (KeyError, IndexError) as e:
         raise Exception(f"Error parsing LLM response: {str(e)}")
+
+

backend/services/transcription.py

@@ -3,9 +3,22 @@
 import platform
 import tempfile
 
+from dotenv import load_dotenv
+
+load_dotenv()
+
+transcription_provider = os.getenv("TRANSCRIPTION_PROVIDER", "local")
+cloud_api_base_url = os.getenv("CLOUD_API_BASE_URL")
+cloud_api_key = os.getenv("CLOUD_API_KEY")
+transcription_model_name = os.getenv("TRANSCRIPTION_MODEL_NAME", "whisper-large-v3")
+
 
 def _init_whisper():
     """Initialize the best Whisper engine based on the available hardware."""
+    # Skip initialization if using cloud provider
+    if transcription_provider == "cloud":
+        return "cloud", None
+
     system = platform.system()
     machine = platform.machine()
 
@@ -37,9 +50,45 @@ def _init_whisper():
 ENGINE, MODEL = _init_whisper()
 
 
+def _transcribe_cloud(audio_path: str) -> str:
+    """Transcribe audio using OpenAI-compatible API"""
+    import requests
+
+    if not cloud_api_key:
+        raise ValueError("CLOUD_API_KEY must be set when using cloud transcription provider")
+
+    # Prepare the endpoint URL
+    base_url = cloud_api_base_url.rstrip("/")
+    if not base_url.endswith("/audio/transcriptions"):
+        endpoint = f"{base_url}/audio/transcriptions"
+    else:
+        endpoint = base_url
+
+    # Prepare the request
+    headers = {
+        "Authorization": f"Bearer {cloud_api_key}"
+    }
+
+    with open(audio_path, "rb") as audio_file:
+        files = {
+            "file": audio_file
+        }
+        data = {
+            "model": transcription_model_name
+        }
+
+        response = requests.post(endpoint, headers=headers, files=files, data=data, timeout=180)
+        response.raise_for_status()
+
+        result = response.json()
+        return result["text"]
+
+
 def transcribe_audio(audio_path: str) -> str:
     """Transcribe audio file to text"""
-    if ENGINE == "mlx":
+    if ENGINE == "cloud":
+        return _transcribe_cloud(audio_path)
+    elif ENGINE == "mlx":
         import mlx_whisper
 
         result = mlx_whisper.transcribe(
@@ -65,3 +114,5 @@ def transcribe_base64(audio_base64: str) -> str:
         return transcribe_audio(temp_path)
     finally:
         os.unlink(temp_path)
+
+