Merge pull request road-core#309 from romartin/AAP-38368-2

AAP-38368: Fixing sample prompt tokens calculations, which may cause further available token calculation issues

Author: tisnik

ols/src/prompts/prompt_generator.py

@@ -13,6 +13,11 @@
 from ols.customize import prompts
 
 
+def restructure_rag_context(text: str, model: str) -> str:
+    """Restructure rag text by appending special characters.."""
+    return restructure_rag_context_post(restructure_rag_context_pre(text, model), model)
+
+
 def restructure_rag_context_pre(text: str, model: str) -> str:
     """Restructure rag text - pre truncation."""
     if ModelFamily.GRANITE in model:

ols/src/query_helpers/docs_summarizer.py

@@ -12,7 +12,11 @@
 from ols.app.models.models import RagChunk, SummarizerResponse
 from ols.constants import RAG_CONTENT_LIMIT, GenericLLMParameters
 from ols.customize import prompts, reranker
-from ols.src.prompts.prompt_generator import GeneratePrompt
+from ols.src.prompts.prompt_generator import (
+    GeneratePrompt,
+    restructure_history,
+    restructure_rag_context,
+)
 from ols.src.query_helpers.query_helper import QueryHelper
 from ols.utils.token_handler import TokenHandler
 
@@ -80,7 +84,12 @@ def _prepare_prompt(
         # Use sample text for context/history to get complete prompt
         # instruction. This is used to calculate available tokens.
         temp_prompt, temp_prompt_input = GeneratePrompt(
-            query, ["sample"], ["ai: sample"], self.system_prompt
+            # Sample prompt's context/history must be re-structured for the given model,
+            # to ensure the further right available token calculation.
+            query,
+            [restructure_rag_context("sample", self.model)],
+            [restructure_history("ai: sample", self.model)],
+            self._system_prompt,
         ).generate_prompt(self.model)
 
         available_tokens = token_handler.calculate_and_check_available_tokens(

ols/utils/token_handler.py

@@ -103,7 +103,7 @@ def calculate_and_check_available_tokens(
             context_window_size - max_tokens_for_response - prompt_token_count
         )
 
-        if available_tokens <= 0:
+        if available_tokens < 0:
             limit = context_window_size - max_tokens_for_response
             raise PromptTooLongError(
                 f"Prompt length {prompt_token_count} exceeds LLM "

tests/unit/prompts/test_prompt_generator.py

@@ -17,8 +17,7 @@
 from ols.src.prompts.prompt_generator import (
     GeneratePrompt,
     restructure_history,
-    restructure_rag_context_post,
-    restructure_rag_context_pre,
+    restructure_rag_context,
 )
 
 model = [GRANITE_13B_CHAT_V2, GPT35_TURBO]
@@ -33,10 +32,7 @@
 
 def _restructure_prompt_input(rag_context, conversation_history, model):
     """Restructure prompt input."""
-    rag_formatted = [
-        restructure_rag_context_post(restructure_rag_context_pre(text, model), model)
-        for text in rag_context
-    ]
+    rag_formatted = [restructure_rag_context(text, model) for text in rag_context]
     history_formatted = [
         restructure_history(history, model) for history in conversation_history
     ]

tests/unit/query_helpers/test_docs_summarizer.py

@@ -1,7 +1,7 @@
 """Unit tests for DocsSummarizer class."""
 
 import logging
-from unittest.mock import ANY, patch
+from unittest.mock import ANY, call, patch
 
 import pytest
 
@@ -121,6 +121,30 @@ def test_summarize_truncation():
     assert summary.history_truncated
 
 
+@patch("ols.utils.token_handler.RAG_SIMILARITY_CUTOFF", 0.4)
+@patch("ols.src.query_helpers.docs_summarizer.LLMChain", new=mock_llm_chain(None))
+def test_prepare_prompt_context():
+    """Basic test for DocsSummarizer to check re-structuring of context for the 'temp' prompt."""
+    summarizer = DocsSummarizer(llm_loader=mock_llm_loader(None))
+    question = "What's the ultimate question with answer 42?"
+    history = ["human: What is Kubernetes?"]
+    rag_index = MockLlamaIndex()
+
+    with patch(
+        "ols.src.query_helpers.docs_summarizer.restructure_rag_context",
+        return_value="patched_history",
+    ) as restructure_rag_context:
+        summarizer.create_response(question, rag_index, history)
+        restructure_rag_context.assert_has_calls([call("sample", ANY)])
+
+    with patch(
+        "ols.src.query_helpers.docs_summarizer.restructure_history",
+        return_value="patched_history",
+    ) as restructure_history:
+        summarizer.create_response(question, rag_index, history)
+        restructure_history.assert_has_calls([call("ai: sample", ANY)])
+
+
 @patch("ols.src.query_helpers.docs_summarizer.LLMChain", new=mock_llm_chain(None))
 def test_summarize_no_reference_content():
     """Basic test for DocsSummarizer using mocked index and query engine."""