Revive Codingo: free-CPU deploy, env-driven Qdrant, SQLite fallback, CPU crash fix

- Dockerfile: CUDA base -> python:3.10-slim CPU; install CPU torch; add spaCy model
- interview_retrieval.py: read Qdrant URL/key from env instead of dead hardcoded cluster
- interview_engine.py: guard unconditional GPU calls that crashed on CPU at import
- app.py: fall back to SQLite when DATABASE_URL is unset; normalise postgres:// scheme
- requirements.txt: drop GPU-only bitsandbytes; add nltk + python-dateutil for resume parser
- add scripts/rebuild_qdrant.py + data/ to rebuild the interview-question vector DB

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.gitignore

@@ -1,3 +1,9 @@
+# Local-only archive of old laptop files (never deploy this)
+everything-old-in-my-laptop/
+
+# Claude / editor working dirs
+.claude/
+
 # Python bytecode
 __pycache__/
 *.py[cod]

Dockerfile

@@ -1,22 +1,38 @@
-# Use an NVIDIA PyTorch container with cuDNN 9.1 support
-FROM nvidia/cuda:12.3.2-cudnn9-runtime-ubuntu22.04
+# CPU-only image for Hugging Face Spaces (free tier).
+# Nothing in the app requires a GPU: the LLM is the Groq API, Whisper runs on
+# CPU, edge-tts is a cloud service, and embeddings use the small MiniLM model.
+FROM python:3.10-slim
 
+ENV OMP_NUM_THREADS=1 \
+    DEBIAN_FRONTEND=noninteractive \
+    PIP_NO_CACHE_DIR=1 \
+    PYTHONUNBUFFERED=1 \
+    # Keep all model/cache downloads inside the writable /tmp dir on Spaces.
+    HF_HOME=/tmp/huggingface \
+    TRANSFORMERS_CACHE=/tmp/huggingface/transformers \
+    HUGGINGFACE_HUB_CACHE=/tmp/huggingface/hub
 
-# Basic setup
-ENV OMP_NUM_THREADS=1
-ENV DEBIAN_FRONTEND=noninteractive
-RUN apt-get update && apt-get install -y \
-    python3 python3-pip ffmpeg git libsndfile1 \
-    # Development tools required to compile native extensions such as llama-cpp-python
-    build-essential cmake libopenblas-dev \
+# System libraries: ffmpeg (audio), libsndfile1 (soundfile/librosa),
+# git (some pip installs), and build tools for any source-only wheels.
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        ffmpeg git libsndfile1 build-essential \
     && rm -rf /var/lib/apt/lists/*
 
-# Set up Python environment
+# Install the CPU build of PyTorch first so the heavy CUDA wheel is never
+# pulled in by transitive dependencies.
+RUN pip install --upgrade pip && \
+    pip install torch==2.1.2 --index-url https://download.pytorch.org/whl/cpu
+
 COPY requirements.txt .
-RUN pip install --upgrade pip && pip install -r requirements.txt
+RUN pip install -r requirements.txt
+
+# Pre-download the small spaCy English model used by the resume parser.
+RUN python -m spacy download en_core_web_sm
 
-# Copy app files
+# Copy the application code.
 COPY . /app
 WORKDIR /app
 
+EXPOSE 7860
+
 CMD ["python3", "app.py"]

README.md

@@ -1,6 +1,38 @@
 ---
 title: Codingo
+emoji: 🤖
+colorFrom: indigo
+colorTo: blue
 sdk: docker
 app_file: app.py
-hardware: gpu
+pinned: false
 ---
+
+# Codingo — AI-Powered Smart Recruitment System
+
+A Flask web app where companies post jobs and candidates apply, then an AI
+interviewer ("LUNA") conducts an automated voice interview and scores the
+candidate.
+
+## Required Space secrets
+
+Set these under **Settings → Variables and secrets**:
+
+| Secret | Purpose |
+| --- | --- |
+| `GROQ_API_KEY` | LLM that generates interview questions and chatbot replies |
+| `QDRANT_API_URL` | URL of the Qdrant cluster holding the interview questions |
+| `QDRANT_API_KEY` | API key for that Qdrant cluster |
+| `DATABASE_URL` | *(optional)* Postgres URL; if unset, the app uses SQLite in /tmp |
+
+## Rebuilding the interview-question vector database
+
+The `interview_questions` Qdrant collection is built from
+`data/shuffled_questions.json` (4233 Q&A pairs, all-MiniLM-L6-v2, 384-dim,
+cosine):
+
+```bash
+export QDRANT_API_URL="https://<cluster>.qdrant.io"
+export QDRANT_API_KEY="<key>"
+python scripts/rebuild_qdrant.py
+```

app.py

@@ -64,8 +64,17 @@ app.config['SESSION_COOKIE_SECURE'] = True
 app.config['REMEMBER_COOKIE_SAMESITE'] = 'None'
 app.config['REMEMBER_COOKIE_SECURE'] = True
 
-# Configure the database connection
-app.config['SQLALCHEMY_DATABASE_URI'] = os.getenv("DATABASE_URL")
+# Configure the database connection.  Use DATABASE_URL when provided (e.g. a
+# hosted Postgres set as a Space secret); otherwise fall back to a local
+# SQLite file in the writable /tmp directory so the app runs with zero
+# database setup.  Note: on Hugging Face the /tmp filesystem is ephemeral, so
+# SQLite data resets when the Space restarts.
+_database_url = os.getenv("DATABASE_URL") or "sqlite:////tmp/codingo.db"
+# SQLAlchemy expects the 'postgresql://' scheme; some providers hand out
+# 'postgres://', which newer SQLAlchemy rejects.  Normalise it.
+if _database_url.startswith("postgres://"):
+    _database_url = _database_url.replace("postgres://", "postgresql://", 1)
+app.config['SQLALCHEMY_DATABASE_URI'] = _database_url
 app.config['SQLALCHEMY_TRACK_MODIFICATIONS'] = False
 
 # Create necessary directories in writable locations

backend/services/interview_engine.py

@@ -30,16 +30,16 @@ try:
 except Exception as e:
     print("⚠️ Qdrant check failed:", e)
 
+# Report GPU availability without assuming a GPU is present.  Calling
+# torch.cuda.get_device_name(0) on a CPU-only host raises and would crash
+# the import (and therefore the whole app), so guard every GPU-only call.
 if torch.cuda.is_available():
     print("🔥 CUDA Available")
-    print(torch.cuda.get_device_name(0))
-    print("cuDNN version:", torch.backends.cudnn.version())
+    print("🧠 GPU:", torch.cuda.get_device_name(0))
+    print("💡 cuDNN version:", torch.backends.cudnn.version())
+    print("💥 cuDNN enabled:", torch.backends.cudnn.is_available())
 else:
-    print("❌ CUDA Not Available")
-print("🔥 CUDA:", torch.cuda.is_available())
-print("🧠 GPU:", torch.cuda.get_device_name(0))
-print("💡 cuDNN version:", torch.backends.cudnn.version())
-print("💥 cuDNN enabled:", torch.backends.cudnn.is_available())
+    print("❌ CUDA Not Available — running on CPU")
 
 
 

backend/services/interview_retrieval.py

@@ -67,15 +67,38 @@ SentenceTransformer = None  # type: ignore
 # ---------------------------------------------------------------------------
 # Qdrant configuration
 #
-# These connection details must not be altered.  They point to the
-# existing Qdrant instance containing interview questions and answers.
-
-if QdrantClient is not None:
-    qdrant_client: QdrantClient | None = QdrantClient(
-        url="https://313b1ceb-057f-4b7b-89f5-7b19a213fe65.us-east-1-0.aws.cloud.qdrant.io:6333",
-        api_key="eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJhY2Nlc3MiOiJtIn0.w13SPZbljbSvt9Ch_0r034QhMFlmEr4ctXqLo2zhxm4",
-    )
+# Connection details are read from the environment so the app can point at
+# whichever Qdrant cluster currently holds the ``interview_questions``
+# collection.  Set QDRANT_API_URL and QDRANT_API_KEY (e.g. as Hugging Face
+# Space secrets, or in a local .env file).  Rebuild the collection from the
+# bundled dataset with ``scripts/rebuild_qdrant.py``.
+import os as _os
+
+_qdrant_url = _os.getenv("QDRANT_API_URL")
+_qdrant_key = _os.getenv("QDRANT_API_KEY")
+
+# Qdrant's REST API listens on :6333; append it if only the bare host is given.
+if _qdrant_url:
+    _qdrant_url = _qdrant_url.rstrip("/")
+    if ".qdrant.io" in _qdrant_url and not _qdrant_url.rsplit(":", 1)[-1].isdigit():
+        _qdrant_url = _qdrant_url + ":6333"
+
+if QdrantClient is not None and _qdrant_url and _qdrant_key:
+    try:
+        qdrant_client: QdrantClient | None = QdrantClient(
+            url=_qdrant_url,
+            api_key=_qdrant_key,
+            check_compatibility=False,
+        )
+    except Exception as _exc:  # pragma: no cover - network/config issues
+        logging.error(f"Failed to initialise Qdrant client: {_exc}")
+        qdrant_client = None
 else:
+    if QdrantClient is not None and not (_qdrant_url and _qdrant_key):
+        logging.warning(
+            "QDRANT_API_URL / QDRANT_API_KEY not set; interview question "
+            "retrieval will fall back to default questions."
+        )
     qdrant_client = None
 
 # Name of the Qdrant collection containing interview Q&A pairs.  Do not

requirements.txt

@@ -22,7 +22,6 @@ inputimeout==1.0.4
 evaluate==0.4.5
 accelerate==0.29.3
 huggingface_hub==0.20.3
-bitsandbytes
 faster-whisper==0.10.0
 edge-tts==6.1.2
 gunicorn
@@ -33,9 +32,9 @@ pydub>=0.25.1
 requests>=2.31.0
 psycopg2-binary
 matplotlib
-bitsandbytes>=0.41.0
 pdfminer
 pdfminer.six
 python-docx
 spacy
-
+nltk
+python-dateutil

scripts/rebuild_qdrant.py

@@ -0,0 +1,121 @@
+"""
+Rebuild the Qdrant ``interview_questions`` collection from the local dataset.
+
+This recreates, byte-for-byte, the vector database the original app used:
+  * collection name : interview_questions
+  * vector size     : 384  (all-MiniLM-L6-v2)
+  * distance        : COSINE
+  * payload         : {"job_role": <lower>, "question": ..., "answer": ...}
+
+The original cluster was deleted after inactivity, but every question is
+preserved in ``data/shuffled_questions.json`` (4233 Q&A pairs, 26 roles),
+which is exactly what populated Qdrant in the first place.
+
+Usage:
+    export QDRANT_API_URL="https://<your-cluster>.qdrant.io:6333"
+    export QDRANT_API_KEY="<your-key>"
+    python scripts/rebuild_qdrant.py
+
+It is safe to re-run: the collection is recreated from scratch each time.
+"""
+
+import json
+import logging
+import os
+import sys
+
+logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")
+
+COLLECTION_NAME = "interview_questions"
+VECTOR_SIZE = 384
+DATA_FILE = os.path.join(
+    os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
+    "data",
+    "shuffled_questions.json",
+)
+
+
+def main() -> int:
+    url = os.getenv("QDRANT_API_URL")
+    key = os.getenv("QDRANT_API_KEY")
+    if not url or not key:
+        logging.error(
+            "Set QDRANT_API_URL and QDRANT_API_KEY environment variables first."
+        )
+        return 1
+
+    # Qdrant cloud URLs need the :6333 REST port; add it if the user pasted
+    # the bare hostname from the dashboard.
+    if url.endswith("/"):
+        url = url[:-1]
+    if ".qdrant.io" in url and not url.rsplit(":", 1)[-1].isdigit():
+        url = url + ":6333"
+
+    from qdrant_client import QdrantClient
+    from qdrant_client.http.models import Distance, PointStruct, VectorParams
+    from sentence_transformers import SentenceTransformer
+
+    logging.info("Loading dataset from %s", DATA_FILE)
+    with open(DATA_FILE, "r", encoding="utf-8") as f:
+        rows = json.load(f)
+    logging.info("Loaded %d Q&A rows", len(rows))
+
+    logging.info("Loading embedding model all-MiniLM-L6-v2 (first run downloads it)")
+    model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
+
+    client = QdrantClient(url=url, api_key=key, check_compatibility=False, timeout=120)
+
+    logging.info("Recreating collection '%s' (size=%d, COSINE)", COLLECTION_NAME, VECTOR_SIZE)
+    client.recreate_collection(
+        collection_name=COLLECTION_NAME,
+        vectors_config=VectorParams(size=VECTOR_SIZE, distance=Distance.COSINE),
+    )
+
+    # Build the points. We embed the QUESTION text, exactly like the original
+    # notebook, and store role/question/answer in the payload.
+    questions, payloads = [], []
+    for item in rows:
+        try:
+            role = item["Job Role"].lower().strip()
+            question = item["Questions"].strip()
+            answer = item["Answers"].strip()
+        except (KeyError, AttributeError):
+            continue
+        if not question:
+            continue
+        questions.append(question)
+        payloads.append({"job_role": role, "question": question, "answer": answer})
+
+    logging.info("Embedding %d questions...", len(questions))
+    vectors = model.encode(questions, batch_size=128, show_progress_bar=True)
+
+    batch_size = 64
+    total = len(questions)
+    for start in range(0, total, batch_size):
+        end = min(start + batch_size, total)
+        points = [
+            PointStruct(id=i, vector=vectors[i].tolist(), payload=payloads[i])
+            for i in range(start, end)
+        ]
+        for attempt in range(1, 4):
+            try:
+                client.upsert(collection_name=COLLECTION_NAME, points=points, wait=True)
+                break
+            except Exception as exc:
+                logging.warning("Batch %d-%d attempt %d failed: %s", start, end, attempt, exc)
+                if attempt == 3:
+                    raise
+        logging.info("Uploaded %d/%d", end, total)
+
+    info = client.get_collection(COLLECTION_NAME)
+    logging.info(
+        "Done. Collection '%s' now has %s points (distance=%s).",
+        COLLECTION_NAME,
+        info.points_count,
+        info.config.params.vectors.distance,
+    )
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())