LCORE-1566: Update llama stack to 0.6.0

pyproject.toml

@@ -28,9 +28,9 @@ dependencies = [
     # Used by authentication/k8s integration
     "kubernetes>=30.1.0",
     # Used to call Llama Stack APIs
-    "llama-stack==0.5.2",
-    "llama-stack-client==0.5.2",
-    "llama-stack-api==0.5.2",
+    "llama-stack==0.6.0",
+    "llama-stack-client==0.6.0",
+    "llama-stack-api==0.6.0",
     # Used by Logger
     "rich>=14.0.0",
     # Used by JWK token auth handler
@@ -160,6 +160,7 @@ llslibdev = [
     "faiss-cpu>=1.11.0",
     "chardet>=5.2.0",
     "psycopg2-binary>=2.9.10",
+    "pypdf>=6.9.2",
     # API scoring: inline::basic
     "requests>=2.33.0",
     # API datasetio: inline::localfs

requirements-build.txt

@@ -16,6 +16,7 @@ flit-core==3.12.0
     #   jinja2
     #   packaging
     #   pathspec
+    #   pypdf
     #   wheel
 hatch-fancy-pypi-readme==25.1.0
     # via
@@ -63,7 +64,7 @@ jinja2==3.1.6
     # via uv-dynamic-versioning
 markupsafe==3.0.3
     # via jinja2
-maturin==1.10.2
+maturin==1.12.6
     # via fastuuid
 packaging==26.0
     # via

requirements.hashes.source.txt

@@ -509,15 +509,15 @@ langdetect==1.0.9 \
 litellm==1.82.6 \
     --hash=sha256:164a3ef3e19f309e3cabc199bef3d2045212712fefdfa25fc7f75884a5b5b205 \
     --hash=sha256:2aa1c2da21fe940c33613aa447119674a3ad4d2ad5eb064e4d5ce5ee42420136
-llama-stack==0.5.2 \
-    --hash=sha256:581fda638088ee029aab20afe3c42ba8f7f6ef21c80bd9ebcae20bb13c3409d3 \
-    --hash=sha256:9334c781e4ded6520aa60c3301a9087e9fb8fdaea8e5f30f8e21d85b17231d8d
-llama-stack-api==0.5.2 \
-    --hash=sha256:6531556dd8bb6555d778360ecfcd850aad7a49a8172b68146995d538e71641f0 \
-    --hash=sha256:a272e4b803fe24a8ba7d22e6d904bf88abd118ba0b6610a20ff5dedb09f38ad7
-llama-stack-client==0.5.2 \
-    --hash=sha256:17c1bbad90f7699da4eb3cae256e8823caa4d2be945512a45c8c6f89ab899f28 \
-    --hash=sha256:473f4d67ac0b243b0fc29555a0203a742615d31bea606b4332d9e2f193f73d6a
+llama-stack==0.6.0 \
+    --hash=sha256:b804830664dc91e54c7225a7a081cb1874c48fc18573569c19fac4a9397e8076 \
+    --hash=sha256:d92711791633f5505a4473ffba3f3e26acb700716fddab5aec419d99e614c802
+llama-stack-api==0.6.0 \
+    --hash=sha256:b99a03aba3659736b6b540c9e5e674b1daac2bf5eeb2a68795113d62b8250672 \
+    --hash=sha256:f0f3a1a6239a5d3b8c7ef02cefdf817c96c6461dcd8a82c1689ac67ec3107270
+llama-stack-client==0.6.0 \
+    --hash=sha256:3290aac36dcafbd1bc0baaf995522e2037f57056672b5a1516af112a4210f3ea \
+    --hash=sha256:7e514a6ffd92f237aceb062dadc4db44e24a3cd9c4ea35e25173d1e0739beb8e
 markupsafe==3.0.3 \
     --hash=sha256:0303439a41979d9e74d18ff5e2dd8c43ed6c6001fd40e5bf2e43f7bd9bbc523f \
     --hash=sha256:068f375c472b3e7acbe2d5318dea141359e6900156b5b2ba06a30b169086b91a \
@@ -637,9 +637,9 @@ multiprocess==0.70.19 \
 nltk==3.9.4 \
     --hash=sha256:ed03bc098a40481310320808b2db712d95d13ca65b27372f8a403949c8b523d0 \
     --hash=sha256:f2fa301c3a12718ce4a0e9305c5675299da5ad9e26068218b69d692fda84828f
-oci==2.169.0 \
-    --hash=sha256:c71bb5143f307791082b3e33cc1545c2490a518cfed85ab1948ef5107c36d30b \
-    --hash=sha256:f3c5fff00b01783b5325ea7b13bf140053ec1e9f41da20bfb9c8a349ee7662fa
+oci==2.170.0 \
+    --hash=sha256:1e205a395e856b6514501d34595dd1e8e261415aaaab2c97f77fd2fc9618dcc3 \
+    --hash=sha256:49adf0ffaf754c07c43194d19d2e11f5e1b602a95bb365bb384d939ff365b1b4
 openai==2.30.0 \
     --hash=sha256:92f7661c990bda4b22a941806c83eabe4896c3094465030dd882a71abe80c885 \
     --hash=sha256:9a5ae616888eb2748ec5e0c5b955a51592e0b201a11f4262db920f2a78c5231d
@@ -874,6 +874,9 @@ pyjwt==2.12.1 \
 pyopenssl==26.0.0 \
     --hash=sha256:df94d28498848b98cc1c0ffb8ef1e71e40210d3b0a8064c9d29571ed2904bf81 \
     --hash=sha256:f293934e52936f2e3413b89c6ce36df66a0b34ae1ea3a053b8c5020ff2f513fc
+pypdf==6.9.2 \
+    --hash=sha256:662cf29bcb419a36a1365232449624ab40b7c2d0cfc28e54f42eeecd1fd7e844 \
+    --hash=sha256:7f850faf2b0d4ab936582c05da32c52214c2b089d61a316627b5bfb5b0dab46c
 pythainlp==5.3.4 \
     --hash=sha256:76744e51e27c895630bafd74f53a1f0aa8782cef2f7f02eebd6427fe8ce8d84d \
     --hash=sha256:e66fd76fb5931834fd4e32ed54337ec62350d7654f187850e4dd4f915e9f624f

requirements.hashes.wheel.txt

@@ -132,11 +132,6 @@ pandas==2.3.3 \
     --hash=sha256:e3a18fae723b808514670a4a0172f9939cdbb095abd5eef1f34cf5ae1b99f424
 peft==0.18.1 \
     --hash=sha256:026817e68c93fcc0569360afa0ee4fb74b06b0a4268240f922bc2bc0a691bcc1
-pillow==12.1.1 \
-    --hash=sha256:58c0a6787ac12753fba61035713f939f33946c970fe48a5756ed1a36c22d2e79 \
-    --hash=sha256:98ab177b9de8751ec5b1dbb7597b45c7edc358f7e16e5764ae93c976c6433f9f \
-    --hash=sha256:dd45829dc58e931ebef6a4b7eb122efef838a8d37848d5ec857a79d4c7e8f543 \
-    --hash=sha256:fcf5fa3497ec5c32843367d1133bc6b6d273e82d2fc86cd6d309cc09e7e457cf
 prometheus-client==0.24.1 \
     --hash=sha256:fe601e041eac55bad8f46da2f3c54f2ab6cd8a8272d9595742c83980e95ed5e4
 prompt-toolkit==3.0.52 \

src/app/endpoints/conversations_v1.py

@@ -3,6 +3,7 @@
 from typing import Any
 
 from fastapi import APIRouter, Depends, HTTPException, Request
+from llama_stack_api import ConversationNotFoundError
 from llama_stack_client import (
     APIConnectionError,
     APIStatusError,
@@ -275,7 +276,8 @@ async def get_conversation_endpoint_handler(  # pylint: disable=too-many-locals,
         ).model_dump()
         raise HTTPException(**response) from e
 
-    except APIStatusError as e:
+    except (APIStatusError, ConversationNotFoundError) as e:
+        # In library mode, ConversationNotFoundError is raised instead of APIStatusError
         logger.error("Conversation not found: %s", e)
         response = NotFoundResponse(
             resource="conversation", resource_id=normalized_conv_id
@@ -382,7 +384,8 @@ async def delete_conversation_endpoint_handler(
         response = ServiceUnavailableResponse(backend_name="Llama Stack", cause=str(e))
         raise HTTPException(**response.model_dump()) from e
 
-    except APIStatusError:
+    except (APIStatusError, ConversationNotFoundError):
+        # In library mode, ConversationNotFoundError is raised instead of APIStatusError
         logger.warning(
             "Conversation %s in LlamaStack not found. Treating as already deleted.",
             normalized_conv_id,
@@ -519,7 +522,8 @@ async def update_conversation_endpoint_handler(
         ).model_dump()
         raise HTTPException(**response) from e
 
-    except APIStatusError as e:
+    except (APIStatusError, ConversationNotFoundError) as e:
+        # In library mode, ConversationNotFoundError is raised instead of APIStatusError
         logger.error("Conversation not found: %s", e)
         response = NotFoundResponse(
             resource="conversation", resource_id=normalized_conv_id

src/app/endpoints/query.py

@@ -46,6 +46,7 @@
 from utils.query import (
     consume_query_tokens,
     handle_known_apistatus_errors,
+    is_context_length_error,
     prepare_input,
     store_query_results,
     update_azure_token,
@@ -303,7 +304,7 @@ async def retrieve_response(
         response = cast(OpenAIResponseObject, response)
 
     except RuntimeError as e:  # library mode wraps 413 into runtime error
-        if "context_length" in str(e).lower():
+        if is_context_length_error(str(e)):
             error_response = PromptTooLongResponse(model=responses_params.model)
             raise HTTPException(**error_response.model_dump()) from e
         raise e

src/app/endpoints/responses.py

@@ -61,6 +61,7 @@
     consume_query_tokens,
     extract_provider_and_model_from_model_id,
     handle_known_apistatus_errors,
+    is_context_length_error,
     store_query_results,
     update_azure_token,
     validate_model_provider_override,
@@ -335,7 +336,7 @@ async def handle_streaming_response(
                 inline_rag_context=inline_rag_context,
             )
         except RuntimeError as e:  # library mode wraps 413 into runtime error
-            if "context_length" in str(e).lower():
+            if is_context_length_error(str(e)):
                 error_response = PromptTooLongResponse(model=api_params.model)
                 raise HTTPException(**error_response.model_dump()) from e
             raise e
@@ -696,7 +697,7 @@ async def handle_non_streaming_response(
                 )
 
         except RuntimeError as e:
-            if "context_length" in str(e).lower():
+            if is_context_length_error(str(e)):
                 error_response = PromptTooLongResponse(model=api_params.model)
                 raise HTTPException(**error_response.model_dump()) from e
             raise e

src/app/endpoints/rlsapi_v1.py

@@ -41,6 +41,7 @@
 from utils.query import (
     extract_provider_and_model_from_model_id,
     handle_known_apistatus_errors,
+    is_context_length_error,
 )
 from utils.responses import (
     build_turn_summary,
@@ -531,8 +532,7 @@ def _map_inference_error_to_http_exception(  # pylint: disable=too-many-return-s
         return HTTPException(**error_response.model_dump())
 
     if isinstance(error, RuntimeError):
-        error_message = str(error).lower()
-        if "context_length" in error_message or "context length" in error_message:
+        if is_context_length_error(str(error)):
             logger.error("Prompt too long for request %s: %s", request_id, error)
             error_response = PromptTooLongResponse(model=model_id)
             return HTTPException(**error_response.model_dump())

src/app/endpoints/streaming_query.py

@@ -85,6 +85,7 @@
     consume_query_tokens,
     extract_provider_and_model_from_model_id,
     handle_known_apistatus_errors,
+    is_context_length_error,
     prepare_input,
     store_query_results,
     update_azure_token,
@@ -354,7 +355,7 @@ async def retrieve_response_generator(
         )
     # Handle know LLS client errors only at stream creation time and shield execution
     except RuntimeError as e:  # library mode wraps 413 into runtime error
-        if "context_length" in str(e).lower():
+        if is_context_length_error(str(e)):
             error_response = PromptTooLongResponse(model=responses_params.model)
             raise HTTPException(**error_response.model_dump()) from e
         raise e
@@ -590,7 +591,7 @@ async def generate_response(
     except RuntimeError as e:  # library mode wraps 413 into runtime error
         error_response = (
             PromptTooLongResponse(model=responses_params.model)
-            if "context_length" in str(e).lower()
+            if is_context_length_error(str(e))
             else InternalServerErrorResponse.generic()
         )
         yield stream_http_error_event(error_response, context.query_request.media_type)
@@ -835,7 +836,7 @@ async def response_generator(  # pylint: disable=too-many-branches,too-many-stat
             )
             error_response = (
                 PromptTooLongResponse(model=context.model_id)
-                if "context_length" in error_message.lower()
+                if is_context_length_error(error_message)
                 else InternalServerErrorResponse.query_failed(error_message)
             )
             yield stream_http_error_event(error_response, media_type)

src/constants.py

@@ -2,7 +2,7 @@
 
 # Minimal and maximal supported Llama Stack version
 MINIMAL_SUPPORTED_LLAMA_STACK_VERSION = "0.2.17"
-MAXIMAL_SUPPORTED_LLAMA_STACK_VERSION = "0.5.2"
+MAXIMAL_SUPPORTED_LLAMA_STACK_VERSION = "0.6.0"
 
 UNABLE_TO_PROCESS_RESPONSE = "Unable to process this request"
 

src/utils/query.py

@@ -51,6 +51,19 @@
 logger = get_logger(__name__)
 
 
+def is_context_length_error(error_message: str) -> bool:
+    """Check if an error message indicates a context length exceeded error.
+
+    Args:
+        error_message: The error message to check.
+
+    Returns:
+        True if the error indicates context length was exceeded.
+    """
+    msg_lower = error_message.lower()
+    return "context_length" in msg_lower or "context length" in msg_lower
+
+
 def store_conversation_into_cache(
     user_id: str,
     conversation_id: str,
@@ -578,10 +591,7 @@ def handle_known_apistatus_errors(
     """
     if error.status_code == 400:
         error_message = getattr(error, "message", str(error))
-        if (
-            "context_length" in error_message.lower()
-            or "context length" in error_message.lower()
-        ):
+        if is_context_length_error(error_message):
             return PromptTooLongResponse(model=model_id)
     elif error.status_code == 429:
         return QuotaExceededResponse.model(model_id)

tests/e2e/features/info.feature

@@ -16,7 +16,7 @@ Feature: Info tests
      When I access REST API endpoint "info" using HTTP GET method
      Then The status code of the response is 200
       And The body of the response has proper name Lightspeed Core Service (LCS) and version 0.5.0
-      And The body of the response has llama-stack version 0.5.2
+      And The body of the response has llama-stack version 0.6.0
 
   @skip-in-library-mode
   Scenario: Check if info endpoint reports error when llama-stack connection is not working