DeepSeek Local Knowledge Base Installer Released: Run Offline at High Speed

The value of an offline installation package lies in minimizing environment setup complexity—but don’t treat it as a black box. After the first installation, I document the software directory, model directory, knowledge base directory, and log location. Knowing these paths makes subsequent migration, backup, and troubleshooting significantly easier.

For validation, perform a disconnection test:

• Import a sample document;
• Ask a question whose answer explicitly appears in that document;
• Then ask another question not covered by the document. The first query should yield a response closely aligned with the source text; the second should explicitly state “no supporting evidence found.” This functional verification is far more meaningful than merely confirming the UI launches successfully.

Over the past two days, I’ve packaged DeepSeek’s integration with a personal knowledge base into a ready-to-run application. You can now launch it instantly on your local machine—no manual code execution required. See details below.

1 Software Functionality Demo

Below is a walkthrough of basic usage: We load a 20-page PDF (deepseek-r1.pdf), enter the search term R1-Zero, and click the Personal Knowledge Search button. The system retrieves relevant passages from the PDF and feeds them as context to a locally deployed deepseek-r1:1.5b model—producing the inference result shown below:

The entire process—search + inference generation—completes in nearly one second. Due to WeChat Official Account limitations on frame count, only the first three frames are captured—no acceleration or speed-up applied. All playback reflects original real-time performance. Tested on a MacBook Pro (M1).

There are two core stages:

1. Knowledge base search,
2. LLM inference.

Compared to vector-database-based knowledge management, this software achieves dramatically faster search speeds. It leverages Whoosh, a highly efficient full-text search engine optimized for large-scale text queries. Whoosh supports fast inverted indexing—ideal for local knowledge bases or embedded search applications.

2 Software Usage Instructions

Starting with this version, we release a fully packaged application—see below. It supports one-click installation: no environment setup, no manual code execution required.

For macOS Users:

• Download the .app bundle.
• Open Terminal, navigate (cd) to the directory containing DeepSeekMine.app.
• Run:

  DeepSeekMine.app/Contents/MacOS/DeepSeekMine
  

For Windows Users:

• Download the .exe installer.
• Double-click to launch—zero external dependencies required.

Prerequisites:

1. Ensure the deepseek-r1:1.5b large language model is already installed on your machine.
2. Verify deepseek-r1:1.5b runs inference correctly locally. If your hardware cannot support it (e.g., due to low memory/GPU), consider switching to a smaller model.
3. Running deepseek-r1:1.5b requires ≥3 GB GPU VRAM. Most newly purchased laptops meet this requirement. For those needing hardware guidance, here’s our minimum recommended configuration:
   • GPU: NVIDIA GTX 1650 (4 GB) or AMD RX 5500 (4 GB)
   • RAM: 16 GB
   • macOS: MacBook Air with M1, M2, or M3 chip (minimum 8 GB unified memory)

👉 To obtain the installation package, reply “Knowledge Base” below.

3 New Features in v0.3

1. Enhanced Chinese Tokenization Accurate segmentation is critical for keyword-based search. Replaced the default jieba tokenizer with PKUSeg, a high-precision segmentation toolkit developed by Peking University. PKUSeg excels at large-scale text processing and delivers superior accuracy.

2. Phrase Matching for Document Retrieval Upgraded from simple keyword matching to flexible phrase matching, enabling better alignment with user input.

   Example:

   • Document A contains: “Xiao Ming’s income & expense ledger…”
   • Pre-update: searching for “Xiao Ming ledger” would fail to match Document A.
   • Post-update: phrase matching uses PKUSeg’s segmentation (“Xiao Ming”, “ledger”) and checks for ordered proximity—i.e., whether both terms appear in sequence within a configurable character span (default: ≤20 characters). Since “Xiao Ming” and “ledger” appear contiguously in Document A, it now matches reliably.

3. Stop Response During DeepSeek Inference Added a Stop Answering button to interrupt ongoing LLM responses mid-generation. If you dislike the current output, simply click the button to halt inference immediately:

   

4. File Icon Display in Personal Knowledge Base Added visual file-type icons to improve document recognition and differentiation:

   

Have questions? Leave comments below—we’ll respond promptly. The software receives regular updates—follow my official posts for announcements.