[WIP] Start of compositional prompting

.vscode/launch.json

@@ -0,0 +1,32 @@
+{
+    // Use IntelliSense to learn about possible attributes.
+    // Hover to view descriptions of existing attributes.
+    // For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
+    "version": "0.2.0",
+    "configurations": [
+        {
+            "name": "Python Debugger: gbnf decoding",
+            "type": "debugpy",
+            "request": "launch",
+            "program": "src/scenicNL/main.py",
+            "console": "integratedTerminal",
+            "args": [
+                "--query_path",
+                "Scenic-CA-AV-Crash/crash_reports/uncategorized/",
+                "--llm_prompt_type",
+                "comp_gbnf",
+                "--count",
+                "1",
+                "-m",
+                "local",
+                "--verbose",
+                "--ignore-cache",
+            ],
+            "env": {
+                "PINECONE_API_KEY":"373c6f4b-61df-40fb-ba26-05e160088c83",
+                "PINECONE_ENVIRONMENT":"gcp-starter",
+                "PINECONE_INDEX":"scenic-programs",
+            }
+        }
+    ]
+}

setup.py

@@ -6,11 +6,10 @@
     packages=find_packages(where="src"),
     package_dir={"": "src"},
     install_requires=[
-        "anthropic",
+        "anthropic==0.19.1",
         "bs4",
-        "carla==0.9.15",
         "scenic==3.0.0b2",
-        "openai>=0.28,<=0.28.1",
+        "openai==1.13.3",
         "pdf2image",
         "pyocr",
         "SQLAlchemy",

src/scenicNL/adapters/anthropic_adapter.py

@@ -1,6 +1,7 @@
 from enum import Enum
 import json
-from typing import cast
+import os
+from typing import List, cast
 
 from anthropic import Anthropic, AI_PROMPT, HUMAN_PROMPT
 import httpx
@@ -12,16 +13,22 @@
 class AnthropicModel(Enum):
     CLAUDE_INSTANT = "claude-instant-1.2"
     CLAUDE_2 = "claude-2.0"
+    CLAUDE_3_STRONG = "claude-3-opus-20240229"
+    CLAUDE_3_MEDIUM = "claude-3-sonnet-20240229"
 
 
 class AnthropicAdapter(ModelAdapter):
     """
     This class servers as a wrapper for the Anthropic API.
     """
-    def __init__(self, model : AnthropicModel):
+    def __init__(self, model : AnthropicModel, use_index : bool = True):
         super().__init__()
         self._model = model
-        self.index = VectorDB(index_name='scenic-programs')
+        if use_index:
+            self.index = VectorDB(index_name='scenic-programs')
+        else:
+            self.index = None
+        self.client = Anthropic(api_key=os.getenv("ANTHROPIC_API_KEY"))
 
     def get_cache_key(
         self, 
@@ -288,6 +295,19 @@ def _format_message(
             raise NotImplementedError(f"Prompt type {prompt_type} was not formatted for Anthropic model {self._model.value}")
 
         return msg
+
+    def predict(
+        self,
+        messages: List[dict],
+    ) -> str:
+        claude_response = self.client.messages.create(
+            messages=messages,
+            model=self._model.value,
+            max_tokens=3600 # Claude's default max tokens
+        )
+
+        return claude_response.content[0].text
+
 
     def _predict(
         self, 

src/scenicNL/adapters/lmql_adapter.py

@@ -2,7 +2,7 @@
 from scenicNL.adapters.api_adapter import Scenic3
 
 import os
-import openai
+from openai import OpenAI
 import json
 from enum import Enum
 from typing import Dict
@@ -22,9 +22,12 @@ class LMQLAdapter(ModelAdapter):
     """
     def __init__(self, model: LMQLModel):
         super().__init__()
-        openai.api_key = os.getenv("OPENAI_API_KEY")
         if os.getenv("OPENAI_ORGANIZATION") and len(os.getenv("OPENAI_ORGANIZATION")) > 0:
-            openai.organization = os.getenv("OPENAI_ORGANIZATION")
+            # TODO: The 'openai.organization' option isn't read in the client API. You will need to pass it when you instantiate the client, e.g. 'OpenAI(organization=os.getenv("OPENAI_ORGANIZATION"))'
+            organization = os.getenv("OPENAI_ORGANIZATION")
+            self.client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"), organization=organization)
+        else:
+            self.client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
         self.PROMPT_PATH = os.path.join(os.curdir, 'src', 'scenicNL', 'adapters', 'prompts')
         self._model = model
 

src/scenicNL/adapters/local_adapter.py

@@ -3,34 +3,15 @@
 
 import requests
 from scenicNL.adapters.model_adapter import ModelAdapter
-from scenicNL.common import LLMPromptType, ModelInput, VectorDB
+from scenicNL.common import LOCAL_MODEL_DEFAULT_PARAMS, LOCAL_MODEL_ENDPOINT, LLMPromptType, ModelInput, VectorDB
+from scenicNL.constraints.gbnf_decoding import CompositionalScenic
 
 
 class LocalModel(Enum):
-    MIXTRAL = "mixtral-8x7b"
+    local = "local"
 
 
 class LocalAdapter(ModelAdapter):
-    ENDPOINT = "http://127.0.0.1:8080/completion"
-    DEFAULT_PARAMS = {
-        "cache_prompt": False,
-        "image_data": [],
-        "mirostat": 0,
-        "mirostat_eta": 0.1,
-        "mirostat_tau": 5,
-        "n_predict": -1,
-        "n_probs": 0,
-        "presence_penalty": 0,
-        "repeat_last_n": 241,
-        "repeat_penalty": 1.18,
-        "slot_id": 0,
-        "stop": ["Question:", "Answer:"],
-        #"stream": False,
-        "tfs_z": 1,
-        "top_k": 40,
-        "top_p": 0.5,
-        "typical_p": 1,
-    }
 
     def __init__(self, model : LocalModel):
         super().__init__()
@@ -68,8 +49,8 @@ def _format_message(
         msg = None
         # TODO: Add more prompt types
 
-        if msg is None:
-            raise NotImplementedError(f"Prompt type {prompt_type} was not formatted for Anthropic model {self._model.value}")
+        # if msg is None:
+        #     raise NotImplementedError(f"Prompt type {prompt_type} was not formatted for Local model {self._model.value}")
 
         return msg
 
@@ -83,13 +64,22 @@ def _predict(
         prompt_type: LLMPromptType,
         verbose: bool
     ) -> str:
+        if prompt_type == LLMPromptType.COMPOSITIONAL_GBNF:
+            program_generator = CompositionalScenic()
+            return program_generator.compositionally_construct_scenic_program(
+                model_input=model_input,
+                temperature=temperature,
+                max_tokens=max_length_tokens,
+                verbose=verbose
+            )
+
         prompt = self._format_message(
             model_input=model_input,
             prompt_type=prompt_type,
             verbose=verbose,
         )
 
-        data = {"prompt": prompt, "temperature": temperature} | self.DEFAULT_PARAMS
+        data = {"prompt": prompt, "temperature": temperature} | LOCAL_MODEL_DEFAULT_PARAMS
         if max_length_tokens > 0:
             data["max_tokens"] = max_length_tokens
 
@@ -99,7 +89,7 @@ def _predict(
         # - Add logic to check the correctness of each partial scenic program
         # - Add logic to synthesize the full scenic program from all of the partial scenic programs
 
-        response = requests.post(self.ENDPOINT, json=data)
+        response = requests.post(LOCAL_MODEL_ENDPOINT, json=data)
         if response.status_code != 200:
             raise ValueError(f"Local model {self._model} returned status code {response.status_code}") 
         response_body = response.json()

src/scenicNL/adapters/model_adapter.py

@@ -156,7 +156,7 @@ def predict_batch(
 
         if verbose:
             print(f"Starting batch prediction using {self.__class__.__name__} " + 
-                  "with {num_workers} workers")
+                  f"with {num_workers} workers")
 
         with Cache(cache_path) as cache:
             processor = self._batch_processor(