Attempt fix for ChromaDB ValueError

app.py

@@ -218,7 +218,7 @@ def to_workflow(request: gr.Request, *args):
             yield value
 
 
-@spaces.GPU(duration=60)
+@spaces.GPU(duration=90)
 def run_workflow_local(*args):
     for value in run_workflow(*args):
         yield value
@@ -264,13 +264,11 @@ with gr.Blocks(
             "local",
             "remote",
         ],
-        value=("local" if torch.cuda.is_available() else "remote"),
+        # Default to remote because it provides a better first impression for most people
+        # value=("local" if torch.cuda.is_available() else "remote"),
+        value="remote",
         label="Compute Mode",
-        info=(
-            "NOTE: remote mode **does not** use ZeroGPU"
-            if torch.cuda.is_available()
-            else "NOTE: local mode requires GPU"
-        ),
+        info="NOTE: remote mode **does not** use ZeroGPU",
         render=False,
     )
 

graph.py

@@ -12,9 +12,6 @@ from retriever import BuildRetriever
 from prompts import query_prompt, generate_prompt, generic_tools_template
 from mods.tool_calling_llm import ToolCallingLLM
 
-# Local modules
-from retriever import BuildRetriever
-
 # For tracing (disabled)
 # os.environ["LANGSMITH_TRACING"] = "true"
 # os.environ["LANGSMITH_PROJECT"] = "R-help-chat"

mods/tool_calling_llm.py

@@ -177,6 +177,10 @@ class ToolCallingLLM(BaseChatModel, ABC):
         # Extract <think>...</think> content and text after </think> for further processing 20250726 jmd
         think_text, post_think = extract_think(response_message.content)
 
+        ## For debugging
+        # print("post_think")
+        # print(post_think)
+
         # Parse output for JSON (support multiple objects separated by commas)
         try:
             parsed_json_results = json.loads(f"[{post_think}]")

retriever.py

@@ -174,9 +174,9 @@ def BuildRetrieverDense(compute_mode: str, top_k=6):
         # Get top k documents
         search_kwargs={"k": top_k},
     )
-    ## Release GPU memory
-    ## https://github.com/langchain-ai/langchain/discussions/10668
-    # torch.cuda.empty_cache()
+    # Fix for ValueError('Could not connect to tenant default_tenant. Are you sure it exists?')
+    # https://github.com/langchain-ai/langchain/issues/26884
+    chromadb.api.client.SharedSystemClient.clear_system_cache()
     return retriever