feat: polish system prompt to support image input handling

Author: lemorage

src/ops/functions/extract_by_llm.rs

@@ -28,13 +28,19 @@ struct Executor {
     value_extractor: base::json_schema::ValueExtractor,
 }
 
-fn get_system_prompt(instructions: &Option<String>, extra_instructions: Option<String>) -> String {
-    let mut message =
-        "You are a helpful assistant that extracts structured information from text. \
-Your task is to analyze the input text and output valid JSON that matches the specified schema. \
-Be precise and only include information that is explicitly stated in the text. \
-Output only the JSON without any additional messages or explanations."
-            .to_string();
+fn get_system_prompt(
+    instructions: &Option<String>,
+    extra_instructions: Option<String>,
+    is_image_input: bool,
+) -> String {
+    let input_type = if is_image_input { "image" } else { "text" };
+    let mut message = format!(
+        "You are a helpful assistant that processes {} inputs to produce structured outputs. \
+Your task is to follow the provided instructions to generate or extract information and output valid JSON matching the specified schema. \
+For generative tasks, provide accurate and relevant responses based on the input. \
+Output only the JSON without additional messages or explanations, unless instructed otherwise.",
+        input_type
+    );
 
     if let Some(custom_instructions) = instructions {
         message.push_str("\n\n");
@@ -54,12 +60,17 @@ impl Executor {
         let client =
             new_llm_generation_client(spec.llm_spec.api_type, spec.llm_spec.address).await?;
         let schema_output = build_json_schema(spec.output_type, client.json_schema_options())?;
+        let is_image_input = args.image.is_some();
         Ok(Self {
             args,
             client,
             model: spec.llm_spec.model,
             output_json_schema: schema_output.schema,
-            system_prompt: get_system_prompt(&spec.instruction, schema_output.extra_instructions),
+            system_prompt: get_system_prompt(
+                &spec.instruction,
+                schema_output.extra_instructions,
+                is_image_input,
+            ),
             value_extractor: schema_output.value_extractor,
         })
     }