Div-AI.
An assistant with somewhere to put the context.
A Telegram assistant prototype that combines local language models with searchable notes, saved conversations, web search, and selected Google tools.
Personal project · Python, Ollama, ChromaDB, Telegram
Saved is a start.
Retrieved is useful.
A controlled replay of a note finding its way into a later answer.
For the portfolio, use warm paper, forest green, and a little chartreuse.
A scripted explanation, not a connected assistant. The prototype combines local models with networked services.
More than the last message.
Recent chat context and retrieved notes help a new question make sense. A short follow-up can use the preceding exchange to form a more useful search query.
Notes are chunked for retrieval. The /save flow summarises a conversation before adding it to memory. Storing information is easy; finding the right piece again is the harder question.
Choosing a path.
The router uses keywords and model classification to choose one path for a message: answer with context, search the web, inspect calendar information, or work with tasks. Telegram updates the answer progressively as the model responds.
Local models, connected tools.
Generation and embeddings use Ollama. Telegram, web search, and Google integrations still use network services. The prototype mixes local inference with connected tools.
Stable chunk IDs prevent identical note chunks being inserted twice. Keeping memory in sync when the source notes change is a separate problem; older chunks can remain.
Things the work
taught me.
Saved is not the same as remembered.
The useful moment is when the right context returns for a new question.
A prototype should show its limits.
Local models still sit beside connected services; the whole assistant is not offline.