|
|
|
|
|
|
by notjulianjaynes
584 days ago
|
|
|
Yeah unfortunately that's the exact model I'm using (Q5 version. What I've been doing is first loading the transcript into the vector database, and then giving it a prompt thats like "summarize the transcript below: <full text of transcript>". This works surprisingly well except for one transcript I had which was of a 3 hour meeting that was per an online calculator about 38,000 tokens. Cutting the text up into 3 parts and pretending each was a seperate meeting* lead to a bunch of hallucinations for some reason. *In theory this shouldn't matter much for my purpose of summarizing city council meetings that follow a predictable format. |
|
|