Doubao PC version is not moral and attacks the open source players who deploy AI knowledge base locally

Written by

Iris Vance

Updated on:June-22nd-2025

At present, in the AI big model market, everyone has recognized that the era of "scenario is king" has arrived. The competition among general AI assistants such as DeepSeek, Doubao, and Yuanbao has become fierce, and the growth space is severely limited. The core driving force for future growth will come from "AI+scenario" applications that can solve specific pain points and provide unique value. Whether it is search, education, creation or social, verticalization and communityization have become the key to breaking the game. Among them, the AI application scenario that is closest to the majority of users and most motivated to practice is: the construction of a personal knowledge base to help themselves and their children grow. And because of the sensitivity of the data, everyone generally hopes to run the AI knowledge base on their own computers. So there are various open source free AI assistant clients + local big models such as "Cherry Studio+DeepSeek local big model" or "dify+Qwen3 local big model" "local AI knowledge base" construction methods, and basically they have formed a track. Tencent ima knowledge base has entered this track before, but the big model is online, and the most important thing is that the knowledge base is online, which makes everyone feel uneasy about some very sensitive data going online. But this time, Doubao PC version quietly started building a "local AI knowledge base", which launched a dimensionality reduction attack on all open source solutions.

Doubao PC version released version v1.49.10 on April 22, 2025. The "Local File Q&A" function was directly launched.

This version of Doubao PC can already use local models for AI Q&A for local files. Files and conversation content will not be uploaded to the server, and the function supports offline use. The large model downloaded by Doubao PC offline belongs to the vectorized version of the Skylark family, namely the Doubao Vector Model. The model capacity exceeds 7GB. This model is designed for vector retrieval scenarios. By generating text embedding vectors with high semantic differentiation, it supports vectorized storage and efficient retrieval of local knowledge bases, and is the core technical foundation for realizing the "local file Q&A" function.

And the model has been accelerated for NVIDIA RTX GPUs, including GeForce RTX 50 series GPUs. Recommended system configuration: RTX GPU with 12GB or more video memory, i5 processor, 16GB memory. (My machine configuration: i9-14900, 32GB memory, RTX 5080 graphics card 16GB video memory).

Doubao PC version provides both local embedding model and local knowledge base together. Haha, it is awesome. It blocks the way for small and medium developers to use open source software to build AI knowledge base. Without further ado, just download the model, upload the knowledge base document, and start working.

The file formats that Doubao PC version allows to upload to the local knowledge base include: DOCX, XLSX, PPTX, PDF, TXT, MD; please note that it does not support uploading image formats, perhaps OCR is not perfect enough. I took a shortcut and directly converted the photo image format into a scanned PDF file, and then uploaded it to the local knowledge base, but it prompted that it could not be recognized, which is also the difficulty of the local large model RAG. I hope that ByteDance will make great efforts in image recognition in the future to make good products and provide users with more practical AI knowledge base products.

In the math problem mentioned in the previous article, "Xuanxuan's shortest path from home to school", Doubao PC version identified and parsed the corresponding knowledge from the knowledge base and answered correctly. At least the knowledge base parsing based on pure text is available, while the recognition and parsing based on multi-modality such as pictures has not yet been perfected.