HumlaChat With Your Meeting History — Private, Cited, and Offline
The problem: your answer is in there somewhere
If you take meeting notes for long enough, you end up with a filing cabinet you can't actually read. The decision you're trying to remember is in a note — you just don't know which one, and scrolling three months of calls to find one sentence is nobody's idea of a good afternoon. Search helps, but only if you remember the exact word someone used. Half the time you don't.
So we built the obvious thing: let you ask.
What shipped
Every note now has a Chat tab, right next to Summary and Transcript. You type a question and Humla goes to work. It doesn't run one search and cross its fingers — it works the question the way you would. It searches, opens the notes that look promising, and if the first search misses, it tries a different angle before answering. You can watch it happen: a live line reads "Searching your notes…", then settles into a receipt like "Searched your notes · Read 2 notes" so you know exactly what it looked at.
Then it answers — and every answer carries citation chips, one per source note, with the title and date. Click a chip and you're in that note, on the day it happened. No "trust me" answers; you can always check the receipt. (For the curious, this is a local flavour of retrieval-augmented generation — the model fetches relevant passages first, then answers grounded in them.)
The private part
Here's the part we care about most: chat can run with nothing leaving your Mac.
Point Humla at Ollama and both halves of the feature — the chat model and the search index — run locally. The chat model is your pick of Gemma or Qwen; the index that powers search is built with EmbeddingGemma, also on-device. Nothing is uploaded, and it works on a plane. The local path isn't a stripped-down demo tier, either — it's the whole product: same search, same citations, same everything.
Prefer a cloud model? Bring your own OpenAI key and only your chat traffic goes to OpenAI, on your own account. There's no Humla server in the middle. And either way, personal chat is free forever — no quotas, no meters. If you chat with OpenAI you pay OpenAI for your own usage; with local models there's nothing to pay at all.
Search that finds it by meaning
The reason chat can find "the pricing discussion from spring" is that notes are indexed in small pieces, by keyword and by meaning. So a question can land on the right stretch of a call from May even if nobody said the word "pricing" — and a passage 40 minutes into an hour-long transcript is just as findable as the first line.
We test it daily in Norwegian and English, including notes that mix both — which, if you take meetings in Norway, is most of them. The models underneath are multilingual, though, spanning a hundred-plus languages, so most languages work fine even if we don't drill them every day.
Scope and sessions
One picker in the composer controls how wide the question goes: this note, its folder, or all notes. Narrow it when you know roughly where to look; open it up when you don't.
And every note can hold more than one conversation. Start a fresh thread when the topic changes, or reopen an old one and pick up where you left off — the history is right there behind a button.
For teams
If your team is on Humla Cloud, the same chat points at your shared workspace instead of just your own notes. Ask about any meeting — whoever actually took it — and get a cited answer. Conversations live in the workspace and are visible to its members, so teammates build on each other's threads instead of asking the same question twice. Your personal chats stay on your Mac and never sync.
Under the hood
For anyone who likes the details: the answer comes from an agentic loop that can take up to six steps, with three tools — search, read a note, and browse. Retrieval is hybrid: SQLite FTS5 (BM25) for keywords plus embedding similarity for meaning, fused with reciprocal rank fusion, over roughly 750-token, paragraph-aware chunks of each note's body, transcript, and summary. Embeddings come from EmbeddingGemma locally or text-embedding-3-small if you're on OpenAI. Notes, the search index, and chat history all live in local SQLite on your disk. There's a fuller write-up on the chat page.
Try it
Chat is in the app now. It's free, it's open source, and it runs on your Mac.