From 70730616ec5b842ce51dd108b1be350786288956 Mon Sep 17 00:00:00 2001
From: Sabiha Khan <87858386+chewwbaka@users.noreply.github.com>
Date: Tue, 24 Mar 2026 09:52:02 +0530
Subject: [PATCH] =?UTF-8?q?fix:=20incorrect=20system=20instruction=20in=20?=
 =?UTF-8?q?llm=20inference=20of=20variable=20extrac=E2=80=A6=20(#201)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* fix: incorrect system instruction in llm inference of variable extraction

* chore: bump pipecat submodule version
---
 .../workflow/pipecat_engine_variable_extractor.py      | 10 +++++++---
 pipecat                                                |  2 +-
 2 files changed, 8 insertions(+), 4 deletions(-)

diff --git a/api/services/workflow/pipecat_engine_variable_extractor.py b/api/services/workflow/pipecat_engine_variable_extractor.py
index 05e198b..3e8bf63 100644
--- a/api/services/workflow/pipecat_engine_variable_extractor.py
+++ b/api/services/workflow/pipecat_engine_variable_extractor.py
@@ -193,15 +193,19 @@ class VariableExtractionManager:
 
         extraction_context = LLMContext()
         extraction_messages = [
-            {"role": "system", "content": system_prompt},
             {"role": "user", "content": user_prompt},
         ]
         extraction_context.set_messages(extraction_messages)
 
         # ------------------------------------------------------------------
-        # Use engine's LLM for out-of-band inference (no pipeline frames)
+        # Use engine's LLM for out-of-band inference (no pipeline frames).
+        # Pass system_prompt via system_instruction so it overrides the
+        # current node's system prompt that build_chat_completion_params
+        # would otherwise prepend.
         # ------------------------------------------------------------------
-        llm_response = await self._engine.llm.run_inference(extraction_context)
+        llm_response = await self._engine.llm.run_inference(
+            extraction_context, system_instruction=system_prompt
+        )
 
         # Get model name for tracing
         model_name = getattr(self._engine.llm, "model_name", "unknown")
diff --git a/pipecat b/pipecat
index 245d6f1..1c804fc 160000
--- a/pipecat
+++ b/pipecat
@@ -1 +1 @@
-Subproject commit 245d6f16843070971f067b302d239e5a2e2346a3
+Subproject commit 1c804fce7dcf65f705471b779600069fded9ffdd