Semantic Recall

What it does

Semantic Recall turns your server's chat history into something you can actually ask. Instead of scrolling or keyword-searching, a member types a question and exceed finds the messages that mean the same thing — even when they use different words — then writes a short answer with links back to the original messages.

Under the hood it is retrieval-augmented search (RAG): every indexed message is stored with an embedding (a numeric fingerprint of its meaning). When someone asks a question, exceed embeds the question, finds the closest messages by similarity, and hands the best matches to Gemini to synthesise a cited answer.

Recall indexes going forward only. Turning it on starts capturing new messages — it does not read or import your existing backlog. The more the server talks after it is enabled, the better recall gets.

Asking a question — ,recall

Any member can ask:

,recall <question>
,askhistory <question>

For example, ,recall what did we decide about the logo colors?

exceed replies with an embed containing a concise, cited answer and a Sources field of up to five jump links, each with a short snippet, so you can open the exact messages the answer came from. If nothing relevant has been indexed, it says so rather than guessing.

Recall is on a short cooldown per user and only works once an admin has enabled it.

Admin setup — ,recallconfig

Members with Manage Server control recall:

,recallconfig on                     turn recall on (start indexing new messages)
,recallconfig off                    turn recall off (keeps the existing index)
,recallconfig all on                 index every channel the bot can read
,recallconfig all off                only index the channels on your list
,recallconfig channel add #chan      add a channel to the index list
,recallconfig channel remove #chan   remove a channel from the index list
,recallconfig status                 show enabled state, mode, channels, and indexed count

There are two indexing modes. With index all on, every readable text channel is captured. With it off, only the channels you add with ,recallconfig channel add are indexed — useful for keeping recall focused on a few discussion channels.

Dashboard

The Semantic Recall page on the dashboard gives you the same controls: an enable toggle, an "index every channel" toggle, a checklist of text channels for the allowlist (disabled while indexing every channel), and a read-only count of how many messages are currently indexed. Changes are saved from the save bar.

Privacy

Recall only indexes the channels you enable — nothing else. It captures normal chat messages going forward; it never backfills history, and it only stores a short excerpt of each message alongside its embedding. Turning recall off stops new messages being indexed. Because recall answers strictly from indexed messages, it can only surface things people already said in channels you chose to index.