Fix tools access by LLM.

Signed-off-by: Aivin V. Solatorio <avsolatorio@gmail.com>

mcp_client.py

@@ -69,6 +69,8 @@ Stay strictly within these boundaries while maintaining a helpful and respectful
 
 
 LLM_MODEL = "claude-3-5-haiku-20241022"
+# What is the military spending of bangladesh in 2014?
+# When a tool is needed for any step, ensure to add the token `TOOL_USE`.
 
 
 class MCPClientWrapper:
@@ -180,13 +182,46 @@ class MCPClientWrapper:
         print(response.content)
         contents = response.content
 
-        while len(contents) > 0:
+        MAX_CALLS = 10
+        auto_calls = 0
+
+        while len(contents) > 0 and auto_calls < MAX_CALLS:
             content = contents.pop(0)
 
             if content.type == "text":
                 result_messages.append({"role": "assistant", "content": content.text})
                 claude_messages.append({"role": "assistant", "content": content.text})
 
+                # if (
+                #     auto_calls < MAX_CALLS
+                #     and "TOOL_USE" in content.text
+                #     and len(contents) == 0
+                # ):
+                #     # Call the LLM automatically.
+                #     claude_messages.append({"role": "user", "content": "ok"})
+                #     try:
+                #         next_response = self.anthropic.messages.create(
+                #             model=LLM_MODEL,
+                #             system=SYSTEM_PROMPT,
+                #             max_tokens=1000,
+                #             messages=claude_messages,
+                #             tools=self.tools,
+                #         )
+                #     except OverloadedError:
+                #         return [
+                #             {
+                #                 "role": "assistant",
+                #                 "content": "The LLM API is overloaded now, try again later...",
+                #             }
+                #         ]
+
+                #     print("next_response", next_response.content)
+
+                #     contents.extend(next_response.content)
+                #     auto_calls += 1
+
+                #     # continue
+
             elif content.type == "tool_use":
                 tool_id = content.id
                 tool_name = content.name
@@ -295,7 +330,9 @@ class MCPClientWrapper:
                         system=SYSTEM_PROMPT,
                         max_tokens=1000,
                         messages=claude_messages,
+                        tools=self.tools,
                     )
+                    auto_calls += 1
                 except OverloadedError:
                     return [
                         {

pyproject.toml

@@ -10,6 +10,7 @@ dependencies = [
     "gradio[mcp]>=5.29.1",
     "httpx>=0.28.1",
     "numpy>=2.2.6",
+    "python-ulid>=3.0.0",
     "scikit-learn>=1.6.1",
     "sentence-transformers>=4.1.0",
 ]

services.py

@@ -71,7 +71,7 @@ class DetailedOutput(SearchOutput):
 def search_relevant_indicators(
     query: str, top_k: int = 1
 ) -> dict[str, list[SearchOutput] | str]:
-    """Search for a shortlist of relevant indicators from the World Development Indicators (WDI) given the query. The search ranking may not be optimal, so the LLM may use this as shortlist and pick the most relevant from the list (if any).
+    """Search for a shortlist of relevant indicators from the World Development Indicators (WDI) given the query. The search ranking may not be optimal, so the LLM may use this as shortlist and pick the most relevant from the list (if any). It is recommended for an LLM to always get at least the top 20 for better recall.
 
     Args:
         query: The search query by the user or one formulated by an LLM based on the user's prompt.

uv.lock

@@ -1711,6 +1711,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/45/58/38b5afbc1a800eeea951b9285d3912613f2603bdf897a4ab0f4bd7f405fc/python_multipart-0.0.20-py3-none-any.whl", hash = "sha256:8a62d3a8335e06589fe01f2a3e178cdcc632f3fbe0d492ad9ee0ec35aab1f104", size = 24546 },
 ]
 
+[[package]]
+name = "python-ulid"
+version = "3.0.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/9a/db/e5e67aeca9c2420cb91f94007f30693cc3628ae9783a565fd33ffb3fbfdd/python_ulid-3.0.0.tar.gz", hash = "sha256:e50296a47dc8209d28629a22fc81ca26c00982c78934bd7766377ba37ea49a9f", size = 28822 }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/63/4e/cc2ba2c0df2589f35a4db8473b8c2ba9bbfc4acdec4a94f1c78934d2350f/python_ulid-3.0.0-py3-none-any.whl", hash = "sha256:e4c4942ff50dbd79167ad01ac725ec58f924b4018025ce22c858bfcff99a5e31", size = 11194 },
+]
+
 [[package]]
 name = "pytz"
 version = "2025.2"
@@ -2140,6 +2149,7 @@ dependencies = [
     { name = "gradio", extra = ["mcp"] },
     { name = "httpx" },
     { name = "numpy" },
+    { name = "python-ulid" },
     { name = "scikit-learn" },
     { name = "sentence-transformers" },
 ]
@@ -2156,6 +2166,7 @@ requires-dist = [
     { name = "gradio", extras = ["mcp"], specifier = ">=5.29.1" },
     { name = "httpx", specifier = ">=0.28.1" },
     { name = "numpy", specifier = ">=2.2.6" },
+    { name = "python-ulid", specifier = ">=3.0.0" },
     { name = "scikit-learn", specifier = ">=1.6.1" },
     { name = "sentence-transformers", specifier = ">=4.1.0" },
 ]