DEV Community: retrovirusretro

The RAG tool that auto-generates Q&A pairs from your documents

retrovirusretro — Wed, 20 May 2026 21:31:46 +0000

Title:

The RAG tool that auto-generates Q&A pairs from your documents
Tags:

ai, docker, ollama, selfhosted
Body:

Most RAG tools split your documents into chunks and embed them. FastGPT does something smarter: it uses an LLM to read your documents and generate question-answer pairs automatically.

27K GitHub stars. Visual workflow builder. Almost no English integration content. Here's the guide.

What is FastGPT?

FastGPT is an LLM-based knowledge platform with two standout features:

1. QA-pair extraction

Instead of naive chunking, FastGPT reads your document with an LLM and extracts pairs like:

Q: What is the return window? → A: 30 days from purchase with original receipt.
Q: Which payment methods are accepted? → A: Visa, Mastercard, PayPal.

These pairs are what gets embedded and retrieved. At query time, question matches question — dramatically more accurate than matching a question against a random document chunk.

Enable it: Dataset → Upload → Processing Mode → QA Split (not Simple Split).

2. Visual workflow builder

FastGPT has a node editor for building branching RAG pipelines without code. Classify intent → route to FAQ or document search → format output. Each step is a configurable node.

⚠️ License Notice

FastGPT uses its own license that prohibits reselling it as a SaaS service to others.

✅ Self-hosted for your own team — OK
✅ As a backend component in your own product — OK
❌ Selling FastGPT as a service to customers — not permitted

If you need commercial freedom, use MaxKB (Apache 2.0) or WeKnora (MIT) instead.

Setup

Clone the repo, copy .env.example to .env, then docker compose up -d. Open localhost:3000 → root / 1234.

FastGPT needs MongoDB (conversation storage) and PostgreSQL with pgvector (vector search). Both are included in the docker-compose.

Full docker-compose: fastgpt-production-stack

Connecting Ollama

FastGPT uses the OpenAI-compatible API that Ollama provides at /v1.

Settings → AI Models → Add:

Provider: OpenAI Compatible
Base URL: http://ollama:11434/v1
API Key: ollama (any non-empty string works)
Model name: llama3

QA-Pair Extraction vs Simple Chunking

This is FastGPT's real advantage. A concrete example:

Your document says: "Returns are accepted within 30 days of the original purchase date, provided the item is in original condition with all packaging."

Simple chunking embeds that sentence as-is. When a user asks "Can I return something after 3 weeks?" the retrieval depends on semantic similarity between the question and that chunk.

QA-pair extraction creates: Q: "What is the return deadline?" → A: "30 days from purchase." Now the question directly matches a question — much higher retrieval confidence.

For knowledge bases where accuracy matters more than speed, this technique consistently outperforms naive chunking.

FastGPT vs the Alternatives

	FastGPT	WeKnora	MaxKB	RAGFlow
QA-pair extraction	✅	❌	❌	❌
Visual workflow	✅	❌	❌	❌
License	Custom*	MIT	Apache 2.0	Apache 2.0
Setup complexity	Medium	Medium	Easy	Hard
PDF table parsing	Good	Basic	Basic	Excellent

Pick FastGPT when: you want QA-pair extraction for maximum accuracy, or you need a visual pipeline builder for complex routing logic.

Production Deployment

FastGPT needs Nginx + SSL for a real domain. The production docker-compose with Nginx config and Let's Encrypt instructions:

→ fastgpt-production-stack

Full Series

This is the last article in the Chinese AI tools series:

Meta repo with Docker Compose + Ollama + n8n for all five:
→ chinese-ai-tools-english-guide

QA-pair extraction is underrated. If you're building a customer support bot or internal knowledge base, try it once — going back to naive chunking feels wrong after.

Chat with your database in plain English — locally, for free

retrovirusretro — Wed, 20 May 2026 21:28:03 +0000

"What were our top 10 customers last quarter by revenue, as a bar chart?"

DB-GPT translates that to SQL, runs it against your database, and renders the chart. No SQL knowledge required. Fully local. MIT licensed. 17K GitHub stars — almost no English content.

What is DB-GPT?

DB-GPT is an open-source framework that puts a natural language interface on top of your databases. You connect PostgreSQL, MySQL, SQLite, or others — then ask questions in plain English. It generates the SQL, executes it, and can visualize results automatically.

Think Metabase meets AI, but fully self-hosted and free.

Supported Databases

PostgreSQL · MySQL · MariaDB · SQLite · ClickHouse · DuckDB · Spark SQL

Setup

Clone the repo, copy .env.example to .env, add your database connection string, then docker compose up -d. Open localhost:5670 → admin / admin → Settings → Database → Add → paste connection string.

Full docker-compose: chinese-ai-tools-english-guide/tools/db-gpt

Example Queries

Once your database is connected:

"Show total sales by month for the last 12 months as a bar chart"
"Which products have inventory below 10 units?"
"Top 5 customers by order value with their email addresses"
"Average order fulfillment time grouped by warehouse"

DB-GPT generates SQL for each, runs it, returns results. Charts render automatically in the UI.

Connecting Ollama

Settings → LLM Provider → Ollama → Base URL: http://ollama:11434/v1 → Model: llama3

For best SQL accuracy use sqlcoder — fine-tuned specifically for SQL generation. Pull it with docker exec -it ollama ollama pull sqlcoder.

DB-GPT vs Vanna.ai

Both let you query databases with natural language:

	DB-GPT	Vanna.ai
License	MIT	MIT
Built-in UI	✅ Full app	Minimal
Charts	✅ Built-in	❌ External
Visual pipeline	✅ AWEL	❌
Self-hosted	✅	✅

DB-GPT if you want a complete self-contained app. Vanna.ai if you're embedding the capability programmatically into your own product.

n8n Automation

A ready-to-import workflow JSON is in the repo (integration/n8n-workflows/db-gpt-query.json). POST {question, db_name} → returns {answer, sql, data}.

Full Guide

→ chinese-ai-tools-english-guide

Previous articles in this series:

Your data never leaves your machine. No API keys, no cloud, no SQL knowledge needed.

The simplest self-hosted RAG you'll ever set up (Apache 2.0, 20K stars)

retrovirusretro — Wed, 20 May 2026 21:24:39 +0000

Most RAG tools make you choose between simplicity and power. MaxKB doesn't try to be powerful — it tries to be simple, and it nails it.

20K+ GitHub stars. Apache 2.0. Almost no English content. Here's the guide.

What is MaxKB?

MaxKB (Max Knowledge Base) is a knowledge base Q&A system by the 1Panel team. It connects to any OpenAI-compatible API — including Ollama — and lets you upload documents, ask questions, and embed a chat widget into any website.

That last part is the killer feature: MaxKB generates a JavaScript snippet that drops a chat widget into any HTML page. One script tag. No iframe, no backend changes needed. Apache 2.0 means you can embed this in commercial products with no restrictions.

Setup

3 commands, under 5 minutes.

Clone the repo, copy .env.example to .env, then docker compose up -d. Open localhost:8081 → admin / admin123 → Settings → Model Provider → Ollama → http://ollama:11434.

Create a knowledge base, upload a PDF, start asking questions. That's the entire setup.

Full docker-compose: maxkb-english-guide

Embed Widget

MaxKB generates a JavaScript snippet — drop it into any HTML page and a chat widget appears bottom-right. No iframe, no backend changes. This is what makes MaxKB unique among RAG tools: it's designed to be embedded.

Go to: Application → your app → Embed → copy the script tag → paste into any HTML page.

API

Works from Python, JavaScript, curl, n8n — anything that speaks HTTP. POST your question to /api/application/{app_id}/chat/completions with a Bearer token. Returns {"content": "the answer"}.

Full Python + JavaScript examples in the repo.

MaxKB vs the Alternatives

	MaxKB	WeKnora	FastGPT	RAGFlow
Setup time	⚡ 3 min	5 min	10 min	15 min
License	Apache 2.0	MIT	Custom*	Apache 2.0
Embed widget	✅	❌	❌	❌
Autonomous agent	❌	✅	✅	❌
PDF table parsing	Basic	Basic	Good	Excellent
Commercial embed	✅	✅	❌	✅

*FastGPT prohibits SaaS resale.

Pick MaxKB when: working knowledge base in 5 minutes, embed widget needed, Apache 2.0 matters.

Pick something else when: complex PDFs → RAGFlow · pipeline builder → FastGPT · multi-hop reasoning → WeKnora

Supported LLM Providers

MaxKB works with any OpenAI-compatible API: Ollama (local, free), OpenAI, Groq, Together, or Anthropic via LiteLLM proxy.

Full Guide

→ github.com/retrovirusretro/maxkb-english-guide

Part of a broader series:
→ chinese-ai-tools-english-guide

MaxKB is the tool I recommend to people who just want RAG working quickly without reading 40 pages of docs.

Tencent just released a RAG framework and nobody's talking about it

retrovirusretro — Wed, 20 May 2026 21:08:48 +0000

In April 2026, Tencent's WeChat team released WeKnora as open source. MIT licensed. Ollama support built-in. Almost zero English content about it.

I spent a few days setting it up, writing the first English integration guide, and comparing it to the alternatives. Here's what I found.

What is WeKnora?

WeKnora is the RAG framework that powers WeChat's Dialog Open Platform — production-tested at a scale most of us will never reach.

At its core it does what every RAG tool does: upload documents, ask questions, get answers grounded in your content.

But it adds two things I haven't seen elsewhere:

1. Autonomous reasoning agent

When you ask a complex question, WeKnora doesn't just search. It plans.

"Compare the pricing strategy in document A with the market analysis in document B" gets decomposed into sub-queries before any retrieval happens. Most RAG tools dump a random mix of chunks into the LLM and hope for the best. WeKnora's agent actually thinks about how to answer before searching.

2. Self-updating knowledge base

Point WeKnora at a URL or folder, set a refresh interval, and it monitors the source and updates the knowledge base automatically when content changes. For internal docs, product catalogs, or anything that evolves — this is genuinely useful.

Setup in 5 Minutes

Two modes. Pick based on what you already have running.

If you already have Ollama running:

"bash
git clone https://github.com/retrovirusretro/weknora-english-guide
cd weknora-english-guide
cp .env.example .env
docker compose up -d"

WeKnora joins your existing Docker network. No duplicate Ollama container.

Fresh install (includes Ollama):
"docker compose -f docker-compose.standalone.yml up -d
docker exec -it weknora-ollama ollama pull llama3"

Open http://localhost:8083 → admin / weknora123 → connect Ollama → upload a PDF → ask a question.

*FastGPT prohibits commercial SaaS resale.

When to pick WeKnora over RAGFlow:

You need the reasoning agent for complex multi-document questions
MIT license matters (embedding in a commercial product)
You want the self-updating KB feature
When to pick RAGFlow instead:

Your PDFs have complex layouts (tables, multi-column, images)
You want a larger English community with more answered questions

n8n Integration
WeKnora exposes a REST API. Connect it to n8n for automation pipelines:
Webhook → WeKnora /api/query → Slack / Email / Notion

A ready-to-import n8n workflow JSON is in the repo:
examples/with-n8n/weknora-query-workflow.json

Import it in n8n → Workflows → Import from file. One click, working webhook.

Why Almost No English Content?
The WeKnora community is on WeChat groups and Zhihu. The maintainers write English READMEs but the tutorial ecosystem never crossed over.

Same story with FastGPT (27K stars), MaxKB (20K stars), DB-GPT (17K stars). Massive Chinese communities, almost nothing in English.

I'm documenting all of them:
→ chinese-ai-tools-english-guide

Full Guide
Everything in this post plus Ollama model selection, production deployment with Nginx + SSL, and the WeKnora vs RAGFlow deep-dive:

→ github.com/retrovirusretro/weknora-english-guide

Have you tried WeKnora? Curious if others run into setup issues I haven't documented yet.

5 Chinese AI tools with 100K+ stars that the West is ignoring

retrovirusretro — Wed, 20 May 2026 21:00:05 +0000

I've been exploring the Chinese open-source AI ecosystem for the past few months. What I found surprised me.

There are tools with 20K, 27K, even 35K GitHub stars — actively maintained, production-ready, MIT or Apache licensed — that have almost zero English community. No Reddit posts. No YouTube tutorials. No Stack Overflow answers.

The docs exist. They're just in Chinese.

Here's what I found, and why it matters.

The 5 Tools

1. WeKnora — Autonomous RAG (Tencent, MIT)

GitHub: Tencent/WeKnora · Released April 2026

WeKnora is the core technology behind WeChat's Dialog Open Platform. It converts raw documents into a queryable knowledge base, but adds something others don't: an autonomous reasoning agent that breaks complex questions into sub-queries before searching.

Ask "Compare pricing across these three competitor docs" — most RAG tools retrieve a random mix of chunks. WeKnora's agent actually plans the retrieval.

Also unique: self-updating knowledge base. Point it at a URL or folder, set a refresh interval, it stays current automatically.

License: MIT → embed in commercial products freely.

2. FastGPT — Visual RAG Workflow Builder (27K ⭐)

GitHub: labring/FastGPT

FastGPT's standout feature is QA-pair extraction: instead of chunking documents blindly, it uses an LLM to generate question-answer pairs from your content. Question matches question at retrieval time — dramatically better accuracy than naive chunking.

It also has a visual node editor for building branching RAG pipelines without code.

License: Custom (self-hosted OK, SaaS resale prohibited).

3. MaxKB — Simplest RAG Setup (20K ⭐, Apache 2.0)

GitHub: 1Panel-dev/MaxKB

MaxKB does one thing well: get a knowledge base running fast and embed it anywhere. It generates a JavaScript widget (one <script> tag) you can drop into any website. No iframe, no complex setup.

Apache 2.0 → commercially embeddable, no restrictions.

("bash
docker compose up -d Done. localhost:8081")

4. DB-GPT — Chat With Your Database (17K ⭐, MIT)

GitHub: eosphoros-ai/DB-GPT

"What were our top 10 customers last quarter by revenue, as a bar chart?"

DB-GPT translates that to SQL, runs it against your PostgreSQL/MySQL/SQLite, and renders the chart. Think Metabase meets AI — but fully local, fully open source.

It supports an AWEL visual pipeline builder for complex multi-step database analysis.

5. RAGFlow — Best PDF Parsing (35K ⭐, Apache 2.0)

GitHub: infiniflow/RAGFlow

Most RAG tools split PDFs by character count. RAGFlow reads the layout: tables stay as tables, headers create structure, multi-column text is handled correctly.

If your documents have complex formatting — financial reports, legal contracts, technical manuals — RAGFlow's chunking quality is noticeably better.

Which One Should You Use?
Need to chat with your DATABASE?
→ DB-GPT
Need the SIMPLEST setup, embeddable widget?
→ MaxKB (Apache 2.0, 3-minute install)
Need a VISUAL workflow builder?
→ FastGPT
Best PDF parsing (tables, images, complex layouts)?
→ RAGFlow
Autonomous reasoning + self-updating KB?
→ WeKnora (newest, MIT)
Shared Infrastructure
All five tools work with Ollama. You don't need an API key for any of them.

I wrote Docker Compose configs for each that plug into a shared Ollama + n8n + Qdrant stack — no duplicate containers, no 5 separate LLMs running.

→ Full English guide with Docker Compose, Ollama integration, and n8n workflows for all five:
github.com/retrovirusretro/chinese-ai-tools-english-guide

Individual deep-dives:

WeKnora English Guide
MaxKB English Guide
FastGPT Production Stack
Why Is There No English Content?
These communities live on WeChat groups, Zhihu, and Bilibili. The maintainers speak English well enough to write a README but the tutorial ecosystem never crossed over.

The pattern reminds me of how Ollama made llama.cpp accessible (40K stars), or how Open-WebUI made Ollama accessible (50K stars). The underlying technology existed. Someone just built the bridge.

These tools are the technology. The bridge is missing.

Have you used any of these? I'm curious what the English-speaking community thinks of them.