A production-focused Retrieval-Augmented Generation (RAG) system for PDF ingestion and grounded Q&A.

This repository includes:

• A root application optimized for fast iteration with Streamlit, FastAPI, and Inngest workflow orchestration.
• A packaged implementation under rag-engine with modular API, core services, and tests.

The goal is practical document intelligence, not notebook demos:

• Multi-provider LLM and embedding support: OpenAI, Gemini, Claude, Ollama, and local fallback paths.
• Durable workflow mode using Inngest step functions.
• Local-first resilience with graceful fallback when remote dependencies are unavailable.
• Source-aware answers with retrieval-backed context.

PDF Upload -> Chunking -> Embeddings -> Qdrant -> Top-K Retrieval -> LLM Answer
               |                               |
               +--------- Inngest Step Functions ----------+

Core design patterns:

• Deterministic IDs for idempotent re-ingestion.
• Provider abstraction via environment configuration.
• Graceful degradation for vector store and model services.
• Preflight diagnostics in UI for developer-friendly troubleshooting.

.
├── main.py                 # FastAPI + Inngest functions (root app)
├── streamlit_app.py        # Streamlit UI with local fallback + preflight checks
├── data_loader.py          # PDF loading, chunking, embeddings
├── vector_db.py            # Qdrant integration with local fallback
├── custom_types.py         # Pydantic models
├── qdrant_storage/         # Embedded/local Qdrant data path
├── uploads/                # Uploaded PDFs
├── doc.md                  # Extended technical walkthrough
└── rag-engine/             # Packaged production structure
    ├── src/rag_engine/
    ├── tests/
    ├── docker/
    └── pyproject.toml

python -m venv .venv
(Set-ExecutionPolicy -Scope Process -ExecutionPolicy RemoteSigned) ; (& .\.venv\Scripts\Activate.ps1)

python -m pip install --upgrade pip
python -m pip install -e .

python -m streamlit run streamlit_app.py

Open: http://127.0.0.1:8501

This mode is enough to ingest and query with local fallback behavior.

Run each service in a separate terminal from repository root.

$env:INNGEST_DEV='1'
python -m uvicorn main:app --host 127.0.0.1 --port 8000

.\.tools\inngest\inngest.exe dev -u http://127.0.0.1:8000/api/inngest --no-discovery

python -m streamlit run streamlit_app.py

Expected endpoints:

• Streamlit: http://127.0.0.1:8501
• FastAPI Inngest endpoint: http://127.0.0.1:8000/api/inngest
• Inngest Dev Server: http://127.0.0.1:8288

The root app reads configuration from .env.

Important variables:

• LLM_PROVIDER: openai | gemini | claude | ollama | local
• EMBED_PROVIDER: openai | gemini | ollama | local
• OLLAMA_BASE_URL
• OLLAMA_MODEL
• OLLAMA_EMBED_MODEL
• INNGEST_ENABLED: true | false
• INNGEST_DEV: 1 | 0
• INNGEST_API_BASE: usually http://127.0.0.1:8288/v1
• INNGEST_EVENT_API_BASE: usually http://127.0.0.1:8288
• QDRANT_PATH: local embedded storage path
• EMBED_DIM: should match embedding model output dimensions

Note: Keep secrets such as API keys in .env and never commit them.

If you want the package-based API/UI implementation:

cd rag-engine
python -m pip install -e .[dev]
python -m uvicorn rag_engine.api.app:app --reload --port 8000

Optional commands (from rag-engine):

pytest
ruff check src tests

From rag-engine directory:

docker compose -f docker/docker-compose.yml up --build

This launches Qdrant and the packaged API service.

• Verify FastAPI is running on port 8000.
• Verify Inngest Dev Server is running on port 8288.
• Confirm .env has INNGEST_API_BASE and INNGEST_EVENT_API_BASE set to localhost endpoints.

• The app will attempt local embedded fallback using QDRANT_PATH.
• Ensure the process can write to qdrant_storage.

• Check provider keys and model names in .env.
• For Ollama, confirm OLLAMA_BASE_URL and model availability.

• Do not commit .env files with real secrets.
• Use local model providers for sensitive workloads when data residency is required.
• Restrict uploaded document handling to trusted sources and controlled environments.

• Multi-provider runtime routing for both embeddings and generation.
• Local fallback mode for degraded infrastructure conditions.
• Deterministic ingestion IDs for idempotency.
• Source-aware answers to improve auditability and trust.

Add your preferred license file (MIT, Apache-2.0, etc.) at repository root.

Built for real-world RAG operations by Adil Shamim.