https://github.com/postor/chatpdf-minimal-demo
a minimal viable project of chatpdf
https://github.com/postor/chatpdf-minimal-demo
Last synced: about 2 months ago
JSON representation
a minimal viable project of chatpdf
- Host: GitHub
- URL: https://github.com/postor/chatpdf-minimal-demo
- Owner: postor
- License: mit
- Created: 2023-03-06T10:43:17.000Z (about 2 years ago)
- Default Branch: main
- Last Pushed: 2024-03-21T14:26:12.000Z (about 1 year ago)
- Last Synced: 2025-03-28T15:10:00.303Z (2 months ago)
- Homepage:
- Size: 964 KB
- Stars: 610
- Watchers: 13
- Forks: 107
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# chatpdf-minimal-demo
[chatpdf](https://www.chatpdf.com/) 的最小实现，和文章对话 | mvp of [chatpdf](https://www.chatpdf.com/)
此项目目的在于研究 chatpdf 实现原理 | this project aims at learn how to build app like chatpdf
## 实现原理 | process flow
- 文章切片到段落 | split articles into pieces
- 通过 OpenAI 的 embedding 接口将每个段落转换为 embedding | convert each piece into embedding with OpenAI
- 将提问的问题转换为 embedding | convert user question into embedding
- 把问题的 embedding 比较所有段落 embedding 得到近似程度并排序 | compare question embedding with all the embeddings of pieces and sort the result
- 把和提问(语义)最接近的一个或几个段落作为上下文，通过 OpenAI 的对话接口得到最终的答案 | use the nearest (meaning) piece(s) as context and ask ChatGPT for the final answer
```
article/pdf question
/ / \ |
piece1 piece2 ........... piece(N) |
| | | |
embedding1 embedding2 ...... embeddingN |
| | | |
--X-------------X---------.....-----------X----------------- question_embedding
| | | |
question question question |
distance1 distance2 distance(N) |
| |
pick nearest piece |
\ /
\ /
construct prompt
|
get answer from ChatGPT
```
## 使用 | usage
```
docker compose up
```