I want to upload chat history and distill them

I want to upload chat history and distill them

If you have chat history with them (photos count too) and want to build their digital presence, start here.

Before you start

  • Only upload records you have the right to use. Modeling a real person’s persona should happen with their consent in advance.
  • Material is stored locally under ./uploads/, never uploaded to any third party. With Ollama you can stay fully offline.
  • Raw material quality decides persona quality — the more chat history, the better; 100+ messages is a good floor.

What you can upload

  • Exported WeChat chat history as .txt
  • CSV
  • Photos, videos

Just drag them into the upload area; no format conversion needed. Beyond that, Feishu, DingTalk, Slack, and email records can be pulled in through the multi-source collectors; if you use SillyTavern, you can import its character cards (PNG v2) directly.

No chat history at all? The guided questionnaire can still build a persona; or try the sample presence first to get a feel for things.

How to do it

1. Create a presence

Fill in the name, your relationship, and how long you’ve been together.

What you should see: the presence is created and you move on to material upload.

2. Upload your material

Drag chat history and photos into the upload area.

What you should see: the files appear in the material list with an uploaded status.

3. Pick a distillation mode

Five named modes: Quick / Standard / Deep / Colleague / Public-figure research.

The mode decides four things: how much chat history is read, how many photos are described, which prompt style is used, and roughly how many model calls it costs. Pick Quick when you want speed, Deep when you want detail; there are dedicated modes for colleagues and public figures. You choose it right on the upload page — no guessing.

What you should see: your chosen mode is confirmed and distillation begins.

4. Wait for the pipeline to finish

Distillation is not “throw the logs at an AI and have it imitate”. It is a multi-stage pipeline:

upload material → person-dimension extraction → structured persona → multi-source fusion → digital presence

What you should see: progress moves forward in real time, stage by stage; at the end you get a finished persona.

What comes out of distillation

The persona has 7 layers: personality traits, speaking style, catchphrases, attachment type, love language, conflict style, moving moments. The goal is not “like them” — it is “is them”.

Finer than those seven layers are the discrimination rules — they turn “like” into “is”:

“When she’s angry she says ‘whatever you say’, but if she says ‘whatever’ — without ‘you say’ — then she’s truly angry.”

Written into “speaking style”, that gets read as “she goes cold when angry”. Written into rules, it stays itself:

trigger → meaning → behavior

Rules can be mined from the chat history automatically: find contrast pairs in the same context that differ by a word or two (“whatever” / “whatever you say”), have the model read out the meaning, then attach the original quote as evidence. Rules without evidence are dropped; rules whose trigger word doesn’t appear in the evidence are dropped too. What gets mined is only a candidate — it enters the prompt only after you confirm it. Once inside, it is a hard constraint, not a style hint.

After generation, check the persona page

Open the persona profile page and look at what the system understood them to be.

What you should see: the seven layers and the discrimination rules are all there. Anything that feels wrong, edit on the spot — start wherever it stops sounding like them.

What’s next

With the persona in hand, you can say the first line — see I want to talk with them. That page covers how to read the emotional states, and why you can see what they remembered.