Packagifying functionsmith

PiperOrigin-RevId: 714443499

experimental/functionsmith/README.md

@@ -18,12 +18,21 @@ limitations under the License.
 
 ## Overview
 
-Functionsmith is a general-purpose problem-solving agent.
+Functionsmith is a general-purpose problem-solving agent using dynamic
+function calling.
 
 USING THIS AGENT IS UNSAFE. It directly runs LLM-produced code, and thus
 should only be used for demonstration purposes.
 
-This agent uses free-form function calling, which means that instead of
+To run with the default task investigating a CSV file with airport data,
+install the package, then run:
+```
+export GOOGLE_API_KEY=my-api-key
+wget https://raw.githubusercontent.com/davidmegginson/ourairports-data/refs/heads/main/airports.csv
+functionsmith_cli
+```
+
+This agent uses dynamic function calling, which means that instead of
 relying on a fixed set of tools predefined in the agent
 [in normal LLM function calling](https://ai.google.dev/gemini-api/docs/function-calling),
 we let the agent itself write with all the functions it needs.
@@ -43,8 +52,8 @@ locally.
 
 The functions are not saved permanently, though this feature can be added.
 
-# TODO(simonf): add a notebook version of the general-purpose agent,
-# as well as an Earth Engine-specific notebook agent.
+TODO(simonf): add a notebook version of the general-purpose agent,
+as well as an Earth Engine-specific notebook agent.
 
 ## Attribution
 

experimental/functionsmith/pyproject.toml

@@ -0,0 +1,49 @@
+[project]
+name = "functionsmith"
+version = "0.1.0"
+description = "Functionsmith is a general-purpose problem-solving agent using dynamic function calling"
+readme = "README.md"
+requires-python = ">=3.9"
+keywords = [
+  "ai",
+  "agent",
+  "llm"
+]
+license = {text = "Apache-2.0"}
+authors = [
+  {name = "Simon Ilyushchenko", email = "simonf@google.com"},
+]
+classifiers = [
+  # Get strings from
+  # http://pypi.python.org/pypi?%3Aaction=list_classifiers
+  "Development Status :: 3 - Alpha",
+  "Intended Audience :: Developers",
+  "Intended Audience :: Science/Research",
+  "License :: OSI Approved :: Apache Software License",
+  "Operating System :: OS Independent",
+  "Programming Language :: Python :: 3 :: Only",
+  "Programming Language :: Python :: 3",
+  "Programming Language :: Python :: 3.9",
+  "Programming Language :: Python :: 3.10",
+  "Programming Language :: Python :: 3.11",
+  "Programming Language :: Python :: 3.12",
+  "Programming Language :: Python :: 3.13",
+  "Programming Language :: Python",
+  "Topic :: Scientific/Engineering :: Artificial Intelligence",
+  "Typing :: Typed",
+]
+dependencies = [
+  "anthropic",
+  "docstring_parser",
+  "google-genai",
+  "openai",
+]
+
+[build-system]
+requires = ['setuptools', 'wheel']
+
+[project.entry-points."console_scripts"]
+functionsmith_cli = "agent:main"
+
+[project.urls]
+Homepage = "https://github.com/google/earthengine-community/tree/master/experimental/functionsmith"

experimental/functionsmith/agent.py → experimental/functionsmith/src/agent.py

@@ -6,7 +6,10 @@
 USING THIS AGENT IS UNSAFE. It directly runs LLM-produced code, and thus
 should only be used for demonstration purposes.
 
-To get started, get a data file:
+To get started, set one of these environment variables to your API key:
+GOOGLE_API_KEY, ANTHROPIC_API_KEY, or OPENAI_API_KEY.
+
+To execute the sample task analyzing airports, get the airports.csv file:
 ```
 wget https://raw.githubusercontent.com/davidmegginson/ourairports-data/refs/heads/main/airports.csv
 ```
@@ -17,9 +20,6 @@
 
 See README.md for more information.
 
-To execute the sample task analyzing airports, get the airtports.csv file from
-https://raw.githubusercontent.com/davidmegginson/ourairports-data/refs/heads/main/airports.csv
-
 The agent will ask the LLM to find "something interesting" about the data
 given its schema. Then the LLM will probably create one or two sets of
 low-level function with tests, then actually analyze the files, then stop
@@ -195,45 +195,51 @@ def task_done(agent_message: str) -> None:
 
 tools = {}
 
-question = task
 
-while True:
-  print(STARS)
-  all_tools = copy.deepcopy(tools)
-  all_tools.update(syscalls)
-  question_with_tools = (
-      question
-      + 'The following functions are available:\n'
-      + '\n'.join([str(x) for x in all_tools.values()])
-  )
-  response = agent.chat(question_with_tools)
-  print(response)
+def main():
+  question = task
 
-  parsed_response = parser.extract_functions(response)
-  if not parsed_response.code and not parsed_response.functions:
-    if parsed_response.error:
-      question = parsed_response.error
-      continue
-    question = input('> ')
-    continue
-
-  tools.update(parsed_response.functions)
-
-  if parsed_response.code:
-    # Concatenate all known source code together.
-    # Functions that were defined in the most recent response will be repeated,
-    # which is okay
-    code_with_tools = (
-        '\n'.join([x.code for x in tools.values()])
-        + '\n'
-        + parsed_response.code
+  while True:
+    print(STARS)
+    all_tools = copy.deepcopy(tools)
+    all_tools.update(syscalls)
+    question_with_tools = (
+        question
+        + 'The following functions are available:\n'
+        + '\n'.join([str(x) for x in all_tools.values()])
     )
+    response = agent.chat(question_with_tools)
+    print(response)
+
+    parsed_response = parser.extract_functions(response)
+    if not parsed_response.code and not parsed_response.functions:
+      if parsed_response.error:
+        question = parsed_response.error
+        continue
+      question = input('> ')
+      continue
 
-    print(STARS)
-    input('HIT ENTER TO RUN THIS CODE')
-    print(STARS)
-    question = code_executor.run_code(code_with_tools, {'task_done': task_done})
-  else:
-    # The response had functions but no code. The agent wanted to define them.
-    # We tell it to go on (that is, to keep writing code).
-    question = 'go on'
+    tools.update(parsed_response.functions)
+
+    if parsed_response.code:
+      # Concatenate all known source code together.
+      # Functions that were defined in the most recent response will be repeated,
+      # which is okay
+      code_with_tools = (
+          '\n'.join([x.code for x in tools.values()])
+          + '\n'
+          + parsed_response.code
+      )
+
+      print(STARS)
+      input('HIT ENTER TO RUN THIS CODE')
+      print(STARS)
+      question = code_executor.run_code(code_with_tools, {'task_done': task_done})
+    else:
+      # The response had functions but no code. The agent wanted to define them.
+      # We tell it to go on (that is, to keep writing code).
+      question = 'go on'
+
+
+if __name__ == '__main__':
+  main()