outlines -> guidance

README.md

@@ -143,7 +143,7 @@ For in-depth descriptions of the above queries, check out our [documentation](ht
 - Easily extendable to [multi-modal usecases](./examples/vqa-ingredient.ipynb) 🖼️
 - Smart parsing optimizes what is passed to external functions 🧠
   - Traverses abstract syntax tree with [sqlglot](https://github.com/tobymao/sqlglot) to minimize LLM function calls 🌳
-- Constrained decoding with [outlines](https://github.com/outlines-dev/outlines) 🚀
+- Constrained decoding with [guidance](https://github.com/guidance-ai/guidance) 🚀
 - LLM function caching, built on [diskcache](https://grantjenks.com/docs/diskcache/) 🔑
 
 ## Quickstart
@@ -246,5 +246,4 @@ Special thanks to those below for inspiring this project. Definitely recommend c
   - As far as I can tell, the first publication to propose unifying model calls within SQL 
   - Served as the inspiration for the [vqa-ingredient.ipynb](./examples/vqa-ingredient.ipynb) example
 - The authors of [Grammar Prompting for Domain-Specific Language Generation with Large Language Models](https://arxiv.org/abs/2305.19234)
-- The maintainers of the [Outlines](https://github.com/outlines-dev/outlines) library for powering the constrained decoding capabilities of BlendSQL
-  - Paper at https://arxiv.org/abs/2307.09702
+- The maintainers of the [Guidance](https://github.com/guidance-ai/guidance) library for powering the constrained decoding capabilities of BlendSQL

benchmark/run.py

@@ -7,9 +7,6 @@
 
 from blendsql import blend
 from blendsql.models import TransformersLLM
-import outlines.caching
-
-outlines.caching.clear_cache()
 
 MODEL = TransformersLLM("HuggingFaceTB/SmolLM-135M", caching=False)
 NUM_ITER_PER_QUERY = 5

blendsql/blend.py

@@ -915,7 +915,7 @@ def blend(
             For example, in `{{LLMMap('convert to date', 'w::listing date')}} <= '1960-12-31'`
             We can infer the output format should look like '1960-12-31' and both:
                 1) Put this string in the `example_outputs` kwarg
-                2) If we have a LocalModel, pass the '\d{4}-\d{2}-\d{2}' pattern to outlines.generate.regex
+                2) If we have a LocalModel, pass the '\d{4}-\d{2}-\d{2}' pattern to guidance
         table_to_title: Optional mapping from table name to title of table.
             Useful for datasets like WikiTableQuestions, where relevant info is stored in table title.
         schema_qualify: Optional bool, determines if we run qualify_columns() from sqlglot

blendsql/ingredients/generate.py

@@ -17,6 +17,9 @@ def generate(model: Model, *args, **kwargs) -> str:
 def generate_openai(
     model: OpenaiLLM, prompt, max_tokens: Optional[int], stop_at: List[str], **kwargs
 ) -> str:
+    """This function only exists because of a bug in guidance
+    https://github.com/guidance-ai/guidance/issues/881
+    """
     client = model.model_obj.engine.client
     return (
         client.chat.completions.create(
@@ -36,7 +39,7 @@ def generate_ollama(
     model: OllamaLLM, prompt, options: Optional[Collection[str]] = None, **kwargs
 ) -> str:
     """Helper function to work with Ollama models,
-    since they're not recognized in the Outlines ecosystem.
+    since they're not recognized natively in the guidance ecosystem.
     """
     if options:
         raise NotImplementedError(

blendsql/models/_model.py

@@ -187,7 +187,7 @@ def _setup(self, *args, **kwargs) -> None:
     @abstractmethod
     def _load_model(self, *args, **kwargs) -> ModelObj:
         """Logic for instantiating the model class goes here.
-        Will most likely be an outlines model object,
+        Will most likely be an guidance model object,
         but in some cases (like OllamaLLM) we make an exception.
         """
         ...

blendsql/models/remote/_openai.py

@@ -52,7 +52,7 @@ class AzureOpenaiLLM(RemoteModel):
         env: Path to directory of .env file, or to the file itself to load as a dotfile.
             Should either contain the variable `OPENAI_API_KEY`,
             or all of `TENANT_ID`, `CLIENT_ID`, `CLIENT_SECRET`
-        config: Optional outlines.models.openai.OpenAIConfig to use in loading model
+        config: Optional dict to use in loading model
         caching: Bool determining whether we access the model's cache
 
     Examples:
@@ -64,14 +64,11 @@ class AzureOpenaiLLM(RemoteModel):
         ```
         ```python
         from blendsql.models import AzureOpenaiLLM
-        from outlines.models.openai import OpenAIConfig
 
         model = AzureOpenaiLLM(
             "gpt-3.5-turbo",
             env="..",
-            config=OpenAIConig(
-                temperature=0.7
-            )
+            config={"temperature": 0.7}
         )
         ```
     """
@@ -124,7 +121,7 @@ class OpenaiLLM(RemoteModel):
         model_name_or_path: Name of the OpenAI model to use
         env: Path to directory of .env file, or to the file itself to load as a dotfile.
             Should contain the variable `OPENAI_API_KEY`
-        config: Optional outlines.models.openai.OpenAIConfig to use in loading model
+        config: Optional argument mapping to use in loading model
         caching: Bool determining whether we access the model's cache
 
     Examples:
@@ -134,14 +131,11 @@ class OpenaiLLM(RemoteModel):
         ```
         ```python
         from blendsql.models import OpenaiLLM
-        from outlines.models.openai import OpenAIConfig
 
-        model = AzureOpenaiLLM(
+        model = OpenaiLLM(
             "gpt-3.5-turbo",
             env="..",
-            config=OpenAIConig(
-                temperature=0.7
-            )
+            config={"temperature": 0.7}
         )
         ```
     """

docs/index.md

@@ -134,7 +134,7 @@ For in-depth descriptions of the above queries, check out our [documentation](ht
 - Easily extendable to [multi-modal usecases](./examples/vqa-ingredient.ipynb) 🖼️
 - Smart parsing optimizes what is passed to external functions 🧠
   - Traverses abstract syntax tree with [sqlglot](https://github.com/tobymao/sqlglot) to minimize LLM function calls 🌳
-- Constrained decoding with [outlines](https://github.com/outlines-dev/outlines) 🚀
+- Constrained decoding with [guidance](https://github.com/guidance-ai/guidance) 🚀
 - LLM function caching, built on [diskcache](https://grantjenks.com/docs/diskcache/) 🔑
 
 <hr>
@@ -161,5 +161,4 @@ Special thanks to those below for inspiring this project. Definitely recommend c
     - As far as I can tell, the first publication to propose unifying model calls within SQL
     - Served as the inspiration for the [vqa-ingredient.ipynb](./examples/vqa-ingredient.ipynb) example
 - The authors of [Grammar Prompting for Domain-Specific Language Generation with Large Language Models](https://arxiv.org/abs/2305.19234)
-- The maintainers of the [Outlines](https://github.com/outlines-dev/outlines) library for powering the constrained decoding capabilities of BlendSQL
-    - Paper at https://arxiv.org/abs/2307.09702
+- The maintainers of the [Guidance](https://github.com/guidance-ai/guidance) library for powering the constrained decoding capabilities of BlendSQL

docs/reference/blenders/blenders.md

@@ -6,7 +6,7 @@ hide:
 
 We use the term "blender" to describe the model which receives the prompts used to perform each ingredient function within a BlendSQL script.
 
-We enable integration with many existing LLMs by building on top of [`outlines` models](https://outlines-dev.github.io/outlines/reference/).
+We enable integration with many existing LLMs by building on top of [`guidance` models](https://github.com/guidance-ai/guidance).
 
 Certain models may be better geared towards some BlendSQL tasks than others, so choose carefully!
 

docs/reference/blenders/ollama.md

@@ -10,7 +10,7 @@ hide:
 
 !!! Note
 
-    We consider Ollama models 'remote', since we're unable to access the underlying logits via outlines. As a result, we can only use Ollama for traditional generation, and not constrained generation (such as via the `options` arg in [LLMQA](../ingredients/LLMQA.md)) 
+    We consider Ollama models 'remote', since we're unable to access the underlying logits. As a result, we can only use Ollama for traditional generation, and not constrained generation (such as via the `options` arg in [LLMQA](../ingredients/LLMQA.md)) 
 
 ## OllamaLLM