ai-pipeline-core 0.1.3__tar.gz → 0.1.4__tar.gz

{ai_pipeline_core-0.1.3 → ai_pipeline_core-0.1.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ai-pipeline-core
-Version: 0.1.3
+Version: 0.1.4
 Summary: Core utilities for AI-powered processing pipelines using prefect
 Project-URL: Homepage, https://github.com/bbarwik/ai-pipeline-core
 Project-URL: Repository, https://github.com/bbarwik/ai-pipeline-core

{ai_pipeline_core-0.1.3 → ai_pipeline_core-0.1.4}/ai_pipeline_core/llm/ai_messages.py

@@ -1,4 +1,5 @@
 import base64
+import hashlib
 import json
 
 from openai.types.chat import (
@@ -59,6 +60,11 @@ class AIMessages(list[AIMessageType]):
                 messages.append(message)
         return messages
 
+    def get_prompt_cache_key(self, system_prompt: str | None = None) -> str:
+        if not system_prompt:
+            system_prompt = ""
+        return hashlib.sha256((system_prompt + json.dumps(self.to_prompt())).encode()).hexdigest()
+
     @staticmethod
     def document_to_prompt(document: Document) -> list[ChatCompletionContentPartParam]:
         """

{ai_pipeline_core-0.1.3 → ai_pipeline_core-0.1.4}/ai_pipeline_core/llm/client.py

@@ -48,15 +48,13 @@ def _process_messages(
         # Use AIMessages.to_prompt() for context
         context_messages = context.to_prompt()
 
-        # Apply caching to context messages
-        for msg in context_messages:
-            if msg.get("role") == "user":
-                # Add cache control to user messages in context
-                msg["cache_control"] = {  # type: ignore
-                    "type": "ephemeral",
-                    "ttl": "120s",  # Cache for 2m
-                }
-            processed_messages.append(msg)
+        # Apply caching to last context message
+        context_messages[-1]["cache_control"] = {  # type: ignore
+            "type": "ephemeral",
+            "ttl": "120s",  # Cache for 2m
+        }
+
+        processed_messages.extend(context_messages)
 
     # Process regular messages without caching
     if messages:
@@ -108,9 +106,14 @@ async def _generate_with_retry(
         **options.to_openai_completion_kwargs(),
     }
 
+    if context:
+        completion_kwargs["prompt_cache_key"] = context.get_prompt_cache_key(options.system_prompt)
+
     for attempt in range(options.retries):
         try:
-            with Laminar.start_as_current_span(model, span_type="LLM", input=messages) as span:
+            with Laminar.start_as_current_span(
+                model, span_type="LLM", input=processed_messages
+            ) as span:
                 response = await _generate(model, processed_messages, completion_kwargs)
                 span.set_attributes(response.get_laminar_metadata())
                 Laminar.set_span_output(response.content)

{ai_pipeline_core-0.1.3 → ai_pipeline_core-0.1.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ai-pipeline-core"
-version = "0.1.3"
+version = "0.1.4"
 description = "Core utilities for AI-powered processing pipelines using prefect"
 readme = "README.md"
 license = {text = "MIT"}
@@ -140,7 +140,7 @@ reportIncompatibleVariableOverride = "error"
 reportMissingParameterType = "warning"
 
 [tool.bumpversion]
-current_version = "0.1.3"
+current_version = "0.1.4"
 commit = true
 tag = true
 tag_name = "v{new_version}"