Code interpreter (#46)

contracts/contracts/OpenAiChatGpt.sol

@@ -91,7 +91,7 @@ contract OpenAiChatGpt {
         stop : "", // null
         temperature : 10, // Example temperature (scaled up, 10 means 1.0), > 20 means null
         topP : 101, // Percentage 0-100, > 100 means null
-        tools : "[{\"type\":\"function\",\"function\":{\"name\":\"web_search\",\"description\":\"Search the internet\",\"parameters\":{\"type\":\"object\",\"properties\":{\"query\":{\"type\":\"string\",\"description\":\"Search query\"}},\"required\":[\"query\"]}}}]",
+        tools : "[{\"type\":\"function\",\"function\":{\"name\":\"web_search\",\"description\":\"Search the internet\",\"parameters\":{\"type\":\"object\",\"properties\":{\"query\":{\"type\":\"string\",\"description\":\"Search query\"}},\"required\":[\"query\"]}}},{\"type\":\"function\",\"function\":{\"name\":\"code_interpreter\",\"description\":\"Evaluates python code in a sandbox environment. The environment resets on every execution. You must send the whole script every time and print your outputs. Script should be pure python code that can be evaluated. It should be in python format NOT markdown. The code should NOT be wrapped in backticks. All python packages including requests, matplotlib, scipy, numpy, pandas, etc are available. Output can only be read from stdout, and stdin. Do not use things like plot.show() as it will not work. print() any output and results so you can capture the output.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"code\":{\"type\":\"string\",\"description\":\"The pure python script to be evaluated. The contents will be in main.py. It should not be in markdown format.\"}},\"required\":[\"code\"]}}}]",
         toolChoice : "auto", // "none" or "auto"
         user : "" // null
         });

contracts/test/OpenAiChatGpt.ts

@@ -38,7 +38,47 @@ describe("OpenAiChatGpt", function () {
       await chatGpt.startChat("Hello");
       // promptId: 0, callbackId: 0
       const openAiConf = await oracle.openAiConfigurations(0)
-      expect(openAiConf.toString()).to.equal("gpt-4-turbo-preview,21,,1000,21,{\"type\":\"text\"},0,,10,101,[{\"type\":\"function\",\"function\":{\"name\":\"web_search\",\"description\":\"Search the internet\",\"parameters\":{\"type\":\"object\",\"properties\":{\"query\":{\"type\":\"string\",\"description\":\"Search query\"}},\"required\":[\"query\"]}}}],auto,")
+      const tools = [
+        {
+          "type": "function",
+          "function": {
+            "name": "web_search",
+            "description": "Search the internet",
+            "parameters": {
+              "type": "object",
+              "properties": {
+                "query": {
+                  "type": "string",
+                  "description": "Search query"
+                }
+              },
+              "required": [
+                "query"
+              ]
+            }
+          }
+        },
+        {
+          "type": "function",
+          "function": {
+            "name": "code_interpreter",
+            "description": "Evaluates python code in a sandbox environment. The environment resets on every execution. You must send the whole script every time and print your outputs. Script should be pure python code that can be evaluated. It should be in python format NOT markdown. The code should NOT be wrapped in backticks. All python packages including requests, matplotlib, scipy, numpy, pandas, etc are available. Output can only be read from stdout, and stdin. Do not use things like plot.show() as it will not work. print() any output and results so you can capture the output.",
+            "parameters": {
+              "type": "object",
+              "properties": {
+                "code": {
+                  "type": "string",
+                  "description": "The pure python script to be evaluated. The contents will be in main.py. It should not be in markdown format."
+                }
+              },
+              "required": [
+                "code"
+              ]
+            }
+          }
+        }
+      ];
+      expect(openAiConf.toString()).to.equal(`gpt-4-turbo-preview,21,,1000,21,{\"type\":\"text\"},0,,10,101,${JSON.stringify(tools)},auto,`);
     });
     it("Oracle can add response", async () => {
       const {chatGpt, oracle, allSigners} = await loadFixture(deploy);

oracles/oracle.py

@@ -9,6 +9,7 @@
 from src.domain.tools.search import web_search_use_case
 from src.domain.knowledge_base import index_knowledge_base_use_case
 from src.domain.knowledge_base import query_knowledge_base_use_case
+from src.domain.tools.code_interpreter import python_interpreter_use_case
 from src.entities import Chat
 from src.entities import FunctionCall
 from src.entities import KnowledgeBaseIndexingRequest
@@ -98,6 +99,12 @@ async def _call_function(function_call: FunctionCall, semaphore: Semaphore):
                     )
                     response = web_search_result.result
                     error_message = web_search_result.error
+                elif function_call.function_type == "code_interpreter":
+                    python_interpreter_result = await python_interpreter_use_case.execute(
+                        formatted_input
+                    )
+                    response = python_interpreter_result.stdout
+                    error_message = python_interpreter_result.stderr
                 else:
                     response = ""
                     error_message = f"Unknown function '{function_call.function_type}'"

oracles/settings.py

@@ -19,8 +19,10 @@
 GCS_BUCKET_NAME = os.getenv("GCS_BUCKET_NAME", "galadriel-assets")
 SERVE_METRICS = os.getenv("SERVE_METRICS", "False").lower() == "true"
 NFT_STORAGE_API_KEY = os.getenv("NFT_STORAGE_API_KEY")
+BEARLY_API_KEY = os.getenv("BEARLY_API_KEY")
 
 KNOWLEDGE_BASE_MAX_SIZE_BYTES = int(
     os.getenv("KNOWLEDGE_BASE_MAX_SIZE_BYTES", 10485760)
 )
 KNOWLEDGE_BASE_CACHE_MAX_SIZE = int(os.getenv("KNOWLEDGE_BASE_CACHE_MAX_SIZE", 100))
+

oracles/src/domain/tools/code_interpreter/entities.py

@@ -0,0 +1,8 @@
+from dataclasses import dataclass
+
+
+@dataclass
+class PythonInterpreterResult:
+    stdout: str
+    stderr: str
+    exit_code: int

oracles/src/domain/tools/code_interpreter/python_interpreter_use_case.py

@@ -0,0 +1,57 @@
+import re
+import base64
+import aiohttp
+import settings
+from typing import Optional
+from src.domain.tools.code_interpreter.entities import PythonInterpreterResult
+
+
+async def _interpret_code(code: str) -> Optional[PythonInterpreterResult]:
+    try:
+        async with aiohttp.ClientSession() as session:
+            async with session.post(
+                "https://exec.bearly.ai/v1/interpreter",
+                headers={
+                    "Authorization": settings.BEARLY_API_KEY,
+                },
+                json={
+                    "fileContents": code,
+                    "inputFiles": [],
+                    "outputDir": "output/",
+                    "outputAsLinks": True,
+                },
+            ) as response:
+                response.raise_for_status()
+                data = await response.json()
+                return PythonInterpreterResult(
+                    stdout=(
+                        base64.b64decode(data["stdoutBasesixtyfour"]).decode()
+                        if data["stdoutBasesixtyfour"]
+                        else ""
+                    ),
+                    stderr=(
+                        base64.b64decode(data["stderrBasesixtyfour"]).decode()
+                        if data["stderrBasesixtyfour"]
+                        else ""
+                    ),
+                    exit_code=data["exitCode"],
+                )
+    except Exception as e:
+        return PythonInterpreterResult(stdout="", stderr=str(e), exit_code=1)
+
+
+async def _strip_markdown_code(md_string: str) -> str:
+    stripped_string = re.sub(r"^`{1,3}.*?\n", "", md_string, flags=re.DOTALL)
+    stripped_string = re.sub(r"`{1,3}$", "", stripped_string)
+    return stripped_string
+
+
+async def execute(code: str) -> Optional[PythonInterpreterResult]:
+    clean_code = await _strip_markdown_code(code)
+    return await _interpret_code(clean_code)
+
+
+if __name__ == "__main__":
+    import asyncio
+
+    print(asyncio.run(execute("print(2 + 2)")))

oracles/src/entities.py

@@ -9,7 +9,7 @@
 from openai.types.chat import ChatCompletion
 from openai.types.chat import ChatCompletionToolParam
 
-ALLOWED_FUNCTION_NAMES = ["image_generation", "web_search"]
+ALLOWED_FUNCTION_NAMES = ["image_generation", "web_search", "code_interpreter"]
 
 OpenAiModelType = Literal["gpt-4-turbo-preview", "gpt-3.5-turbo-1106"]
 

oracles/template.env

@@ -13,6 +13,7 @@ ORACLE_ADDRESS="0x"
 ORACLE_ABI_PATH="../arbitrum-contracts/artifacts/contracts/Oracle.sol/Oracle.json"
 
 GCS_BUCKET_NAME="galadriel-assets"
+BEARLY_API_KEY=""
 SERVE_METRICS="False"
 NFT_STORAGE_API_KEY=""