ai_summary Goes Fully Deterministic: Your Instruction Is the Prompt, Byte for Byte
The ai_summary API no longer edits your prompt on the way through. What you send in instruction is what the model reads, byte for byte, so a prompt you tuned once behaves the same on every call.
01 Nothing between you and the model
Previously the platform assembled the final prompt from your input. It could reorder sections, relabel them, append its own instructions or shorten what you sent. That made prompt tuning guesswork: a change that worked in one call could land differently in the next, because the text the model actually saw was not the text you wrote.
Now there is no assembly step. The bytes you send are the bytes the model receives.
Your instruction was one input among several. The platform decided the order, the labels and what else to add, and the result could shift between calls.
Your instruction is the prompt. Nothing is reordered, relabelled, appended or shortened, so the same input gives the same behaviour every time.
02 Retrieval is yours to control
Because the prompt is now entirely under your control, so is retrieval. Run embed_and_search, take the top results you want, and place them in the prompt exactly where you want them — before the question, after it, grouped, trimmed, however your application needs them.
03 Existing integrations keep working
- query and context still work. The older parameters remain supported, so integrations built against them do not need changing.
- Prompts up to 100,000 characters. Enough room to place a substantial set of retrieved documents directly in the instruction.
04 All AI integrations updated
LangChain, LlamaIndex, Haystack, the MCP server and Laravel Scout all ship the new answer behaviour. Update with pip install -U or composer update.