Commit Graph

40 Commits

Author SHA1 Message Date
Jorge Manrubia 3d8437bd5f Truncate summaries to keep within API limits
https://fizzy.37signals.com/5986089/cards/1739
2025-09-15 16:14:29 +02:00
Jorge Manrubia eabce24415 Switch to gpt-5-chat
Cheaper and good enough

https://3.basecamp.com/2914079/buckets/37331921/todos/9041305517
2025-09-12 16:16:54 +02:00
Jason Zimdars 3814c17d49 Be defensive about unwanted elements it sometimes inserts 2025-09-04 12:52:01 -05:00
Jason Zimdars 53f06d2554 Adjust summary, style in newspaper format 2025-09-02 22:14:13 -05:00
Jorge Manrubia b6b4617eec Remove unused constant 2025-09-02 17:24:22 +02:00
Jorge Manrubia 80e37bef36 Separate method with two outputs into two 2025-08-27 14:32:13 +02:00
Jorge Manrubia 744faf12c0 Extract "response cost" object to unify calculations around LLM usage costs 2025-08-27 13:18:41 +02:00
Jorge Manrubia 0eea4f003c Remove unnecessary guards 2025-08-27 09:41:47 +02:00
Jorge Manrubia 853a75a54f Rename "cost_microcents" => "cost_in_microcents" 2025-08-27 09:38:23 +02:00
Jason Zimdars 5cf119cbd1 Adjust prompt to force proper link construction; recognize people milestones and events 2025-08-18 10:00:16 -05:00
Jason Zimdars acfe89ca52 Update prompt to render links and use more casual language 2025-08-14 14:32:53 -05:00
Jason Zimdars e8dc14fda7 Simplify prompt, adjust styles to move away from bullet format 2025-08-11 18:03:48 -05:00
Jorge Manrubia 8addd4bd50 Remove comment 2025-08-11 12:42:31 +02:00
Jason Zimdars 9362b711d8 Adjust prompt 2025-08-05 16:02:11 -05:00
Jason Zimdars b30a329a1a Surface some fun user activity 2025-07-31 14:30:08 -05:00
Jason Zimdars 7d55069d88 Adjust prompt 2025-07-31 10:30:00 -05:00
Jason Zimdars b795a43d5c Simplify the prompt, try a top 5 points format 2025-07-30 18:14:37 -05:00
Jason Zimdars 0f3aa31943 Bold names, ensure summaries generate correctly structured collection links 2025-07-29 11:57:23 -05:00
Jason Zimdars 3479129721 Update prompt 2025-07-25 10:07:49 -05:00
Jorge Manrubia 018c746d2a Make the prompt smarter about the current context to make better calls 2025-07-24 16:48:03 +02:00
Jorge Manrubia 8bb7783215 Extract "prompt generation" logic to models for events, comments and cards
So that we can reuse in other scenarios
2025-07-24 16:48:03 +02:00
Jorge Manrubia 4d3e7bf0af Fix: llm model 2025-07-24 16:48:01 +02:00
Jorge Manrubia f3fa324f09 Let summaries know about collections 2025-07-23 16:54:39 +02:00
Jorge Manrubia c082a1b1f5 More tweaks 2025-07-23 13:15:49 +02:00
Jorge Manrubia c8c297a75e Fix model selection (I had broken it!) 2025-07-23 13:10:47 +02:00
Jorge Manrubia 49518692bb Some prompt refinements 2025-07-23 13:04:41 +02:00
Jorge Manrubia fc1ff7bf5e Fix tests 2025-07-23 12:21:20 +02:00
Jorge Manrubia 90ddadea5c Format 2025-07-23 12:04:18 +02:00
Jorge Manrubia f32e89c277 Format 2025-07-23 11:43:10 +02:00
Jorge Manrubia c3ae05833e Make models selectable 2025-07-23 11:02:34 +02:00
Jorge Manrubia 5322ce2afb Add comment's context for each card too 2025-07-22 22:43:01 +02:00
Jorge Manrubia e1ab4ec13c Use 4o for now, it's just better for summarizing and conversation
It's more expensive, going with the best and we'll tune later
2025-07-22 22:34:12 +02:00
Jorge Manrubia 662f6dffe8 Use gpt-4.1 for now, several prompt refinements 2025-07-22 21:32:10 +02:00
Jorge Manrubia fd0484176c Format 2025-07-22 18:59:10 +02:00
Jorge Manrubia 13b20d43c8 Refine prompt 2025-07-22 18:13:44 +02:00
Jorge Manrubia 0397327442 Avoid links to comments for now 2025-07-22 14:40:59 +02:00
Jorge Manrubia 6d26195bc7 Add a debug prompt endpoint so that we can play with the prompts live 2025-07-22 14:33:34 +02:00
Jorge Manrubia 89a9e7f32c Tweak prompt for trying to avoid liks like "card 1234" 2025-07-22 12:28:05 +02:00
Jorge Manrubia 053ae76312 Tweak prompt 2025-07-22 12:06:56 +02:00
Jorge Manrubia 7cd459f2b8 AI Summaries: basic abstractions, functional version 2025-07-22 11:57:39 +02:00