pingcap / tidb.ai

This is an GraphRAG - Knowledge Graph based and out-of-the-box conversational search tool that leverages the vector storage capabilities of TiDB Serverless. It provides a seamless way to embed a powerful question-answering (QA) bot directly on your website, requiring only a simple copy-and-paste of a JavaScript snippet. Demo: https://tidb.ai

TiDB.AI

Introduction

A [WIP] conversational search RAG (Retrieval-Augmented Generation) app based on TiDB Serverless Vector Storage, providing a out-of-the-box and embeddable QA robot experience based on your knowledge on official and documentation sites.

Live Demo: TiDB.AI

With this tool, you can achieve:

Features

1. Perplexity-style Conversational Search page: Our platform features an advanced built-in website crawler, designed to elevate your browsing experience. This crawler effortlessly navigates official and documentation sites, ensuring comprehensive coverage and streamlined search processes through sitemap URL scraping.

2. Embeddable JavaScript Snippet: Integrate our conversational search window effortlessly into your website by copying and embedding a simple JavaScript code snippet. This widget, typically placed at the bottom right corner of your site, facilitates instant responses to product-related queries.

Quick Start [Working in Progress]

To deploy the application in a self-hosted environment, run the following command:

curl https://tidb.cloud/install.sh | sh

then:

ticloud create-app --template rag

Deployment [TODO]

For deploying the application to production, there are many options available:

• Deploy with Docker
• Deploy to Vercel

Tech Stack

• Next.js – Framework
• TypeScript – Language
• Tailwind – CSS
• shadcn/ui - Design
• TiDB – Database to store chat history, vector, json, and analytic
• Kysely - SQL query builder
• NextAuth.js – Auth
• Vercel – Deployments
• LlamaIndex - RAG framework

License

TiDB.AI is open-source under the Apache License, Version 2.0. You can find it here.

tidb.ai's Issues

Change the default mode to dark

Move the /authentication to /settings as a tab

chat model: support Claude3 in aws bedrock

Skip searching while the questions is not related to TiDB

https://tidb.ai/c/VlARoUrD2E8E

Move "Index Tasks" to a tab inside "Indexes" in Admin panel

milestone 1: eating our own dog food

Background

As we develop Vector Search in TiDB Serverless, we'll build an app to test how easy it is to use. This app will provide answers to TiDB usage questions on our official websites using our in-progress vector storage in TiDB Serverless.

Todo list

• core rag logic
  • data source management
    • upload pdf/markdown/csv etc.
    • crawl sitemap.xml of a domain
  • ui/ux
    • conversational search
      • chat history mgmt after login
    • embeddable js
• system settings
  • overview: statistics of chats & docs
  • basic info settings
    • logo
    • site name
    • search title & subtitle
    • example questions, max to 4
    • footer links
    • GitHub / Discord etc. social media links top-right
  • oauth configurations
    • GitHub
    • Gmail
  • (pre|post| prompt settings
  • rag loader & spliter configruations: chunk size, overlap etc.
• Use llamaindex-ts as RAG engine
• add widget on the bottom-right of (www|ask).pingcap.com to answer questions about tidb usage / use cases

/blog & /docs: use icon-dark.svg when in light mode, icon-light.svg when dark mode

favicon

Support multiple LLM & Embedding method

As there are many stronger LLM that chatgpt4 created, we should have ability to add more embedding method.

The footer links is hidden on Pad

Optimize the URL structure

https://tidb.ai/conversations/HWgaKxMYl4to -> https://tidb.ai/c/HWgaKxMYl4to

Shorter is better! 🤞

Add social.discord support

Style bugs on mobile screen

Finish /c pages

https://tidb.ai/c, need to add some apis maybe:

• GET /c, to list all conversations, with paging
• DELETE /c/, with a simple alert

Graph editor doesn't render very well

https://tidb.ai/c/4wo1YnvkJtaR

The answer seems too simple

https://tidb.ai/c/dSHuudxqQXUY

bug: dustbin icon isn't visible when the question is too long

Style error of https://tidb.ai/sources

Move static values of RAG to /settings

• Social Media of RAG
• Title & Sub Title
• Example Questions( just leave 5 there in /settgings)

Move the social icon from bottom to top-right in small screen

Add provider prefix to user id

Add a /docs module like shadcn

Code Ref:
https://github.com/shadcn-ui/ui/tree/main/apps/www

Style:
https://ui.shadcn.com/docs

menu items: Docs(https://tidb.ai/docs), Demo(https://tidb.ai)
add social media icon/github, mode switcher to top-right, and remove the same one on the bottom-left.

Entrance of /docs should be put at the bottom-center(sticky to bottom with a suitable margin)

Move "Import Tasks" to somewhere in "Overview"

Maybe a tab

Support cost statistics

Calculate the cost of RAG in calling various third-party services, like:

• Serverless Function (Vercel)
• Storage
• Embedding
• Rerank
• LLM Completion
• Database Query (TiDB Serverless)

Add Google reCAPTCHA config in /settings/security

Always add a related(next) questions list at the bottom ,just behind the last question

3 related(next) questions according to the previous context.

Enable robots.txt & sitemap in www/, a nextra dir

Bug: Two active tab

Support embedding & reranker of Jina AI

Description

After the PR run-llama/LlamaIndexTS#734 merged. LlamaindexTS will support embedding and reranker of Jina AI.

Tasks

• #78
• Create a new "Index" with embedding & reranker of Jina AI for test

Support metadata extraction and filter for more accurate query

Tasks

• #98
• Support metadata extraction from HTML label
• Add post filters to filter the ANN search result by metadata #122
• Add config page for the metadata extraction rule
• Experiment: Embedding text combined with the metadata extracted by document

chat model: support https://chatglm.cn/

Support Bitdeer as provider

Description

Bitdeer is a cloud computing platform that provides computing power for cryptocurrency mining.

In the AI age, Bitdeer also provides multiple managed AI models (For example, llama2, mistral, etc.) API services.

Tasks

• #97
• #107
• #114
• API support streaming mode

Move these hard-code titles to /settings page

Examples Questions wlll also be used in embeddable js scripts

Wrong citation

[^2]: [TiDB Official Documentation](https://tidb.net/blog/ce4e4dd6)

Enable "Share" button

It doesn't work currently.

[The Most Important Feature] Move embeddable js script copy/paste button to /settings

We should provide a copy button of js scrip somewhere in /settings page, to allow user to copy js and then paste it to his/her blogs, webiste to provide the bottom-right chatbot service.

• Move code from @shczhen to pingcap/tidb.ai
• Add JavaScript tab in /settings
• Make this widget configurable, e.g.: icon, button-text, dialog title, subtitle, example questions etc.
  

feature: multiple indexing config

• support multiple indexing config (splitter, embedding, ...) for same document
• UI to compare conversation quality

Langfuse Integration

Langfuse does not provide TypeScript version of Llamaindex support (reference: langfuse/langfuse-js#200), so we need to realize it ourself.

• Work around
• Add LangfuseContext
• Add CallbackHandler, reference https://github.com/langfuse/langfuse-js/blob/main/langfuse-langchain/src/callback.ts
• Add @observe decorator

Remove colors from /docs, we use black/white style

[draft] milestone 2: private beta

Background

For our next milestone, we'll focus on making our app easier to deploy and use:

Product Refinement:

• Dedicate efforts to polishing the product, ensuring it meets user expectations in usability.

Deployment:

• Add support for deploying on Vercel(maybe and Cloudflare).
• Make it possible to deploy locally using tools like npx for quick testing.

Content Updates:

• Create a welcoming landing page for new visitors.
• Write usage docs/api docs to help users get started and integrate.

Todo list

• #80
• product
  • api & api token mgmt
  • more LLM support(maybe need to re-split/re-index/re-embedding the content of docs)
  • support more data source: word etc.
  • Role-based access control & mgmt(user, admin)
• deployment
  • manually deployment(with npx etc.)
  • docker
  • Vercel
  • Cloudflare(optional)
  • fly.io(optional)
  • LLM/tokens usage statistic
• security
  • cross-domain white list
• content
  • landing page
    • /
    • /showcases
    • /docs/get-started
    • /docs/api
  • README.md

Deployment guide

• For self-hosted
• Vercel

The sidebar footer is hidden on small screen

Fix /home style on small screen

For title&subtitle, we can make the width shorter and wrap the text.

For docs in footer, we can make it one line per document, like this:

Line1: Docs
Line2: Another Link
Line3: Linkeeee4

how to connect Ollama?

no idea...

[Draft] Implement an incremental crawler

src/core/interface.ts

export namespace rag {
  export interface Content<ContentMetadata> {
    content: string[];
    digest: string;
+  lastModifiedAt: Date;
    metadata: ContentMetadata;
  }

  export type ImportSourceTaskResult = {
    enqueue?: Array<{ type: string, url: string }>
    content?: {
      buffer: Buffer
      mime: string
    }
+   incrementalState?: unknown
  }

  export abstract class ImportSourceTaskProcessor<Options> extends Base<Options> {
    abstract support (taskType: string, url: string): boolean;
    abstract process (task: { url: string }): Promise<ImportSourceTaskResult>

+   abstract supportIncremental (taskType: string, url: string): boolean;
+   abstract processIncremental (previousState: unknown): Promise<ImportSourceTaskResult>
  }
}

src/core/db/importSource.ts

export interface ImportSource {
  created_at: Date;
  filter: string | null;
  filter_runtime: string | null;
  id: string;
  type: string;
  url: string;
+ incremental_state: JSON | null;
+ last_scheduled_at: Date | null;
}

[Never Close] Image Storage

We paste images into this GitHub Issue, then retrieve the image URL for use in displaying images in the README.

Add footer links to /settings to be configurable too:

Add footer links to /settings to be configurable too: