lake-flow-pipeline 0.1.1

Data Lake pipelines for Vector DB, RAG & AI. Ingest, process, embed, and semantic search.

LakeFlow Backend

FastAPI backend and data pipelines for LakeFlow: ingest, staging, processing, embedding, and semantic search.

---

Overview

• API: FastAPI app (lakeflow.main:app) — auth, search, embed, pipeline trigger, Qdrant proxy, system.
• Data Lake: Layered zones under LAKEFLOW_DATA_BASE_PATH: 000_inbox → 100_raw → 200_staging → 300_processed → 400_embeddings → 500_catalog.
• Vector store: Qdrant (default collection lakeflow_chunks). Embeddings via sentence-transformers (e.g. all-MiniLM-L6-v2).

---

Requirements

• Python ≥ 3.10
• Qdrant (e.g. Docker: docker compose up -d qdrant)
• See requirements.txt for Python dependencies

---

Install & run

With Docker (from the LakeFlow repo root where docker-compose.yml is):

docker compose up --build
# API: http://localhost:8011

Local dev (from repo root, go to lakeflow):

cd lakeflow
python3 -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt
pip install -e .
# Create/copy .env (repo root or lakeflow) with LAKEFLOW_DATA_BASE_PATH, QDRANT_HOST, etc.
python -m uvicorn lakeflow.main:app --reload --port 8011

• If you get bad interpreter (venv points to wrong Python): remove .venv, run python3 -m venv .venv again then pip install -r requirements.txt and pip install -e ..

• If you get Address already in use (port 8011 in use): free the port then restart the server — lsof -ti :8011 | xargs kill -9

• Swagger: http://localhost:8011/docs

• ReDoc: http://localhost:8011/redoc

• Embed API: docs/API_EMBED.md — POST /search/embed

---

Pipeline steps (CLI)

Run from the lakeflow directory (with venv activated and LAKEFLOW_DATA_BASE_PATH set in .env or environment).

Step	Command	Output
0 – Inbox → Raw	python -m lakeflow.scripts.step0_inbox	Hash, dedup, catalog
1 – Staging	python -m lakeflow.scripts.step1_raw	pdf_profile.json, validation.json
2 – Processed	python -m lakeflow.scripts.step2_staging	clean_text.txt, chunks.json, tables.json
3 – Embeddings	python -m lakeflow.scripts.step3_processed_files	embeddings.npy, chunks_meta.json
4 – Qdrant	python -m lakeflow.scripts.step3_processed_qdrant	Points in Qdrant

Or use the Streamlit UI (Pipeline Runner) when LAKEFLOW_MODE=DEV.

---

Main APIs

• POST /auth/login – Demo login (e.g. admin / admin123), returns JWT.
• POST /search/embed – Body {"text": "..."} → vector, embedding, dim.
• POST /search/semantic – Body {"query": "...", "top_k": 5, "qdrant_url": "...", "collection_name": "..."}.
• POST /search/qa – RAG-style Q&A (semantic search + LLM). Optional.
• POST /pipeline/run – Run a pipeline step (auth required).
• GET/POST /qdrant/ – Qdrant collections and points (proxy).

---

Design notes

• Idempotent pipelines; deterministic UUIDs for Qdrant.
• SQLite without WAL (NAS-friendly).
• No full-file load for large files; streaming where applicable.

---

License

Same as the root repository.