feat: short prompt for fine-tuned converter models

Fine-tuned models were trained with the short 'Convert recipe to
Cooklang:' prompt (recipe-pack finetune crate) but production sent them
the full ~1.4k-token instruction prompt - a train/inference mismatch
that also wastes tokens on every call. Models with the ft: prefix now
get the exact prompt they were trained with; base models keep the full
instruction prompt.

Author: dubadub

src/converters/mod.rs

@@ -12,7 +12,9 @@ pub use evaluator::{inject_evaluation, COOKLANG_EVALUATOR_PROMPT};
 pub use google::GoogleConverter;
 pub use ollama::OllamaConverter;
 pub use open_ai::OpenAiConverter;
-pub use prompt::{inject_recipe, COOKLANG_CONVERTER_PROMPT};
+pub use prompt::{
+    inject_recipe, prompt_for_model, COOKLANG_CONVERTER_PROMPT, FINETUNED_CONVERTER_PREFIX,
+};
 
 use async_trait::async_trait;
 use serde::Serialize;

src/converters/open_ai.rs

@@ -1,4 +1,4 @@
-use super::{inject_recipe, ConversionMetadata, ConversionResult, Converter, TokenUsage};
+use super::{prompt_for_model, ConversionMetadata, ConversionResult, Converter, TokenUsage};
 use crate::config::ProviderConfig;
 use async_trait::async_trait;
 use log::debug;
@@ -93,7 +93,7 @@ impl Converter for OpenAiConverter {
             .json(&json!({
                 "model": self.model,
                 "messages": [
-                    {"role": "user", "content": inject_recipe(content)}
+                    {"role": "user", "content": prompt_for_model(&self.model, content)}
                 ],
                 "temperature": self.temperature,
                 "max_tokens": self.max_tokens,

src/converters/prompt.rs

@@ -28,6 +28,24 @@ pub fn inject_recipe(recipe_content: &str) -> String {
         .replace("{{LANGUAGE}}", &language)
 }
 
+/// The user prompt fine-tuned converter models were trained with
+/// (see recipe-pack's finetune crate).
+pub const FINETUNED_CONVERTER_PREFIX: &str = "Convert recipe to Cooklang:\n\n";
+
+/// Builds the converter prompt for the given model.
+///
+/// Fine-tuned models (`ft:` prefix) get the short prompt they were trained
+/// with — the instruction set is baked into their weights, and sending the
+/// full rulebook both mismatches their training distribution and costs
+/// ~1.4k extra tokens per call. Base models get the full instruction prompt.
+pub fn prompt_for_model(model: &str, recipe_content: &str) -> String {
+    if model.starts_with("ft:") {
+        format!("{}{}", FINETUNED_CONVERTER_PREFIX, recipe_content)
+    } else {
+        inject_recipe(recipe_content)
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -44,6 +62,22 @@ mod tests {
         assert!(COOKLANG_CONVERTER_PROMPT.contains("timer"));
     }
 
+    #[test]
+    fn test_prompt_for_model_finetuned_uses_short_prompt() {
+        let p = prompt_for_model(
+            "ft:gpt-4.1-mini-2025-04-14:personal::abc",
+            "2 eggs\n\nBoil.",
+        );
+        assert_eq!(p, "Convert recipe to Cooklang:\n\n2 eggs\n\nBoil.");
+    }
+
+    #[test]
+    fn test_prompt_for_model_base_uses_full_prompt() {
+        let p = prompt_for_model("gpt-4.1-mini", "2 eggs\n\nBoil.");
+        assert!(p.contains("Cooklang syntax rules"));
+        assert!(p.contains("2 eggs\n\nBoil."));
+    }
+
     #[test]
     fn test_prompt_contains_examples() {
         // Verify the prompt includes examples