Helpdesks don't fail because agents are slow. They fail because search is keyword-based — it matches words, not meaning.
This engine indexes 25,000 real consumer financial complaints and finds semantically similar cases using transformer embeddings. Search "my bank charged me twice" and it surfaces relevant past complaints — even if none of them contain those exact words.
Built to production standards: A/B testing framework to compare embedding models live, MLflow experiment tracking, Prometheus monitoring, and drift detection on the query embedding space.
Dataset: 25,000 sampled from 500K+ CFPB complaints
Search quality: 0.645 similarity score on mortgage queries
Throughput: 37 sentences/second (MiniLM)
Models compared: MiniLM vs MPNet via 90/10 traffic split
🔗 API: https://semantic-search-v1-2.onrender.com
📖 Interactive Docs: https://semantic-search-v1-2.onrender.com/docs
⚠️ Hosted on Render free tier. First request after inactivity triggers a ~10 minute re-index. If the API returns"status": "indexing", wait and retry.
Demo queries to try:
?query=mortgage+late+fees?query=credit+card+fraud?query=student+loan+payment+problems?query=bank+account+closed+without+notice
docker pull spectra2204/semantic-search:v1.1
docker run -p 8000:8000 spectra2204/semantic-search:v1.1
⚠️ First run takes ~10 minutes to index 25,000 complaints into ChromaDB.
Watch the logs — once you see✅ Ingestion complete, the API is ready.
Visit http://localhost:8000/docs to explore the API interactively.
| Component | Technology | Why |
|---|---|---|
| Embeddings | SentenceTransformers (MiniLM) | 6x faster than MPNet with similar accuracy |
| Vector Store | ChromaDB | Persistent vector storage with metadata filtering |
| API | FastAPI | Best-in-class Python backend with auto docs |
| Experiment Tracking | MLflow | Model registry + UI for comparing runs |
| Monitoring | Prometheus | Per-model latency and similarity score tracking |
| A/B Routing | MD5 hash % 100 | Deterministic — same query always hits same model |
Semantic search using MiniLM. Returns ranked complaints matching query intent.
{
"query": "mortgage late fees",
"results": [{
"score": 0.645,
"product": "Mortgage",
"issue": "Loan servicing, payments, escrow account",
"text": "my mortgage company charge me late fee for the last nine months..."
}]
}Same search with 90/10 traffic split — 10% routed to MPNet for model
comparison. Response includes model_version field showing which model
served the request.
Returns API status and whether 25,000 complaints are indexed and ready.
Prometheus metrics — search latency, request counts, per-model similarity scores.
MiniLM over MPNet — Both models returned similar accuracy, but MiniLM processes queries 6x faster, making it the right default for production latency.
Re-index on startup over persistent volume — Cloud volume setup adds infrastructure complexity with minimal benefit for a 25K dataset. Re-indexing on cold start is rare and keeps the stack simple.
Hash-based routing over random — MD5 hashing ensures the same query always routes to the same model, making A/B comparison meaningful and reproducible.
Project-2-Semantic-Search/
├── data/
│ └── complaints_clean.csv
├── notebooks/
│ ├── 01_eda.ipynb
│ └── 02_embeddings.ipynb
├── src/
│ ├── api.py
│ └── startup.py
├── Dockerfile
├── .dockerignore
└── requirements.txt
- Project 1 — Fraud Detection API: XGBoost + FastAPI + Docker + Railway
🔗 https://fraud-ml-pipeline.onrender.com/docs
📦 github.com/Spectraa28/fraud-ml-pipeline