Azure OpenAI RAG that keeps shedding side quests

What it does

entaoai is an integration accelerator that demonstrates ChatGPT-like experiences over private data using Azure OpenAI and LangChain. It provides a web interface and API to upload documents, index them, and query them through chat or QnA interfaces. The codebase is TypeScript with Python API components, designed to run on Azure Functions and Web Apps.

The interesting bit

The README reads like an archaeological dig: the project has aggressively refactored itself, spinning off SEC analysis, financial copilots, SQL agents, and advanced Prompt Flow variants into separate repositories. What remains is a distilled RAG core that notably refuses to lock you into one vector backend, offering Pinecone, Redis, and Azure Cognitive Search interchangeably. It also bakes in LLM-assisted evaluation flows—groundedness, coherence, F1—via Azure Prompt Flow, which is more rigor than your average demo.

Key highlights

• Swappable vector stores: supports Pinecone, Redis, and Azure Cognitive Search for retrieval.
• Dual API backends: choose between Azure Functions or Azure Prompt Flow managed endpoints.
• Evaluation suite: includes Prompt Flow evaluations for groundedness, coherence, similarity, and F1 score.
• Streaming chat and function calling: supports real-time responses and optional tool use for weather, stocks, and Bing search.
• Recent diet: as of March 2024, peripheral features were moved to other repos, leaving a leaner core focused on chat, QnA, upload, and admin.

Caveats

• The README is essentially a chronological changelog, so understanding the current architecture requires reading between the lines.
• The project history includes repeated breaking changes—OpenAI SDK upgrades, endpoint renames, and feature removals—so stability is a moving target.
• Some promised updates, like Prompt Flow support for Redis and Cognitive Search in autonomous flows, were noted as pending but remain unverified in the provided text.

Verdict

A useful reference if you need a working Azure RAG prototype with flexible backends and can tolerate tracing through a dense commit history. Look elsewhere if you need a polished, stable product rather than a wiring diagram come to life.