llm-ie 0.4.4__tar.gz → 0.4.5__tar.gz

{llm_ie-0.4.4 → llm_ie-0.4.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: llm-ie
-Version: 0.4.4
+Version: 0.4.5
 Summary: An LLM-powered tool that transforms everyday language into robust information extraction pipelines.
 License: MIT
 Author: Enshuo (David) Hsu
@@ -41,6 +41,10 @@ An LLM-powered tool that transforms everyday language into robust information ex
     - Support for LiteLLM.
 - [v0.4.1](https://github.com/daviden1013/llm-ie/releases/tag/v0.4.1) (Jan 25, 2025): Added filters, table view, and some new features to visualization tool (make sure to update [ie-viz](https://github.com/daviden1013/ie-viz)).
 - [v0.4.3](https://github.com/daviden1013/llm-ie/releases/tag/v0.4.3) (Feb 7, 2025): Added Azure OpenAI support. 
+- [v0.4.5](https://github.com/daviden1013/llm-ie/releases/tag/v0.4.5) (Feb 16, 2025): 
+    - Added option to adjust number of context sentences in sentence-based extractors.
+    - Added support for OpenAI reasoning models ("o" series).
+
 
 ## Table of Contents
 - [Overview](#overview)
@@ -340,6 +344,14 @@ from llm_ie.engines import OpenAIInferenceEngine
 inference_engine = OpenAIInferenceEngine(model="gpt-4o-mini")
 ```
 
+For reasoning models ("o" series), use the `reasoning_model=True` flag. The `max_completion_tokens` will be used instead of the `max_tokens`. `temperature` will be ignored.
+
+```python
+from llm_ie.engines import OpenAIInferenceEngine
+
+inference_engine = OpenAIInferenceEngine(model="o1-mini", reasoning_model=True)
+```
+
 #### <img src=doc_asset/readme_img/Azure_icon.png width=32 /> Azure OpenAI API
 In bash, save the endpoint name and API key to environmental variables `AZURE_OPENAI_ENDPOINT` and `AZURE_OPENAI_API_KEY`.
 ```
@@ -356,6 +368,14 @@ from llm_ie.engines import AzureOpenAIInferenceEngine
 inference_engine = AzureOpenAIInferenceEngine(model="gpt-4o-mini")
 ```
 
+For reasoning models ("o" series), use the `reasoning_model=True` flag. The `max_completion_tokens` will be used instead of the `max_tokens`. `temperature` will be ignored. 
+
+```python
+from llm_ie.engines import AzureOpenAIInferenceEngine
+
+inference_engine = AzureOpenAIInferenceEngine(model="o1-mini", reasoning_model=True)
+```
+
 #### 🤗 huggingface_hub
 The ```model``` can be a model id hosted on the Hugging Face Hub or a URL to a deployed Inference Endpoint. Refer to the [Inference Client](https://huggingface.co/docs/huggingface_hub/en/package_reference/inference_client) documentation for more details. 
 
@@ -783,7 +803,7 @@ frames = extractor.extract_frames(text_content=text, entity_key="Diagnosis", str
 
 The ```SentenceFrameExtractor``` instructs the LLM to extract sentence by sentence. The reason is to ensure the accuracy of frame spans. It also prevents LLMs from overseeing sections/ sentences. Empirically, this extractor results in better recall than the ```BasicFrameExtractor``` in complex tasks. 
 
-For concurrent extraction (recommended), the `async/ await` feature is used to speed up inferencing. The `concurrent_batch_size` sets the batch size of sentences to be processed in cocurrent.
+For concurrent extraction (recommended), the `async/await` feature is used to speed up inferencing. The `concurrent_batch_size` sets the batch size of sentences to be processed in cocurrent.
 
 ```python
 from llm_ie.extractors import SentenceFrameExtractor
@@ -792,15 +812,29 @@ extractor = SentenceFrameExtractor(inference_engine, prompt_temp)
 frames = extractor.extract_frames(text_content=text, entity_key="Diagnosis", case_sensitive=False, fuzzy_match=True, concurrent=True, concurrent_batch_size=32)
 ```
 
-The ```multi_turn``` parameter specifies multi-turn conversation for prompting. If True, sentences and LLM outputs will be appended to the input message and carry-over. If False, only the current sentence is prompted. For LLM inference engines that supports prompt cache (e.g., Llama.Cpp, Ollama), use multi-turn conversation prompting can better utilize the KV caching and results in faster inferencing. But for vLLM with [Automatic Prefix Caching (APC)](https://docs.vllm.ai/en/latest/automatic_prefix_caching/apc.html), multi-turn conversation is not necessary.
+The `context_sentences` sets number of sentences before and after the sentence of interest to provide additional context. When `context_sentences=2`, 2 sentences before and 2 sentences after are included in the user prompt as context. When `context_sentences="all"`, the entire document is included as context. When `context_sentences=0`, no context is provided and LLM will only extract based on the current sentence of interest.
 
 ```python
 from llm_ie.extractors import SentenceFrameExtractor
 
-extractor = SentenceFrameExtractor(inference_engine, prompt_temp)
-frames = extractor.extract_frames(text_content=text, entity_key="Diagnosis", multi_turn=False, case_sensitive=False, fuzzy_match=True, stream=True)
+extractor = SentenceFrameExtractor(inference_engine=inference_engine, 
+                                   prompt_template=prompt_temp, 
+                                   context_sentences=2)
+frames = extractor.extract_frames(text_content=text, entity_key="Diagnosis", case_sensitive=False, fuzzy_match=True, stream=True)
 ```
 
+For the sentence:
+
+*The patient has a history of hypertension, hyperlipidemia, and Type 2 diabetes mellitus.*
+
+The context is "previous sentence 2" "previous sentence 1" "the sentence of interest" "proceeding sentence 1" "proceeding sentence 2":
+
+*Emily Brown, MD (Cardiology), Dr. Michael Green, MD (Pulmonology)
+
+*#### Reason for Admission*
+*John Doe, a 49-year-old male, was admitted to the hospital with complaints of chest pain, shortness of breath, and dizziness. The patient has a history of hypertension, hyperlipidemia, and Type 2 diabetes mellitus. #### History of Present Illness*
+*The patient reported that the chest pain started two days prior to admission. The pain was described as a pressure-like sensation in the central chest, radiating to the left arm and jaw.*
+
 </details>
 
 <details>

{llm_ie-0.4.4 → llm_ie-0.4.5}/README.md

@@ -24,6 +24,10 @@ An LLM-powered tool that transforms everyday language into robust information ex
     - Support for LiteLLM.
 - [v0.4.1](https://github.com/daviden1013/llm-ie/releases/tag/v0.4.1) (Jan 25, 2025): Added filters, table view, and some new features to visualization tool (make sure to update [ie-viz](https://github.com/daviden1013/ie-viz)).
 - [v0.4.3](https://github.com/daviden1013/llm-ie/releases/tag/v0.4.3) (Feb 7, 2025): Added Azure OpenAI support. 
+- [v0.4.5](https://github.com/daviden1013/llm-ie/releases/tag/v0.4.5) (Feb 16, 2025): 
+    - Added option to adjust number of context sentences in sentence-based extractors.
+    - Added support for OpenAI reasoning models ("o" series).
+
 
 ## Table of Contents
 - [Overview](#overview)
@@ -323,6 +327,14 @@ from llm_ie.engines import OpenAIInferenceEngine
 inference_engine = OpenAIInferenceEngine(model="gpt-4o-mini")
 ```
 
+For reasoning models ("o" series), use the `reasoning_model=True` flag. The `max_completion_tokens` will be used instead of the `max_tokens`. `temperature` will be ignored.
+
+```python
+from llm_ie.engines import OpenAIInferenceEngine
+
+inference_engine = OpenAIInferenceEngine(model="o1-mini", reasoning_model=True)
+```
+
 #### <img src=doc_asset/readme_img/Azure_icon.png width=32 /> Azure OpenAI API
 In bash, save the endpoint name and API key to environmental variables `AZURE_OPENAI_ENDPOINT` and `AZURE_OPENAI_API_KEY`.
 ```
@@ -339,6 +351,14 @@ from llm_ie.engines import AzureOpenAIInferenceEngine
 inference_engine = AzureOpenAIInferenceEngine(model="gpt-4o-mini")
 ```
 
+For reasoning models ("o" series), use the `reasoning_model=True` flag. The `max_completion_tokens` will be used instead of the `max_tokens`. `temperature` will be ignored. 
+
+```python
+from llm_ie.engines import AzureOpenAIInferenceEngine
+
+inference_engine = AzureOpenAIInferenceEngine(model="o1-mini", reasoning_model=True)
+```
+
 #### 🤗 huggingface_hub
 The ```model``` can be a model id hosted on the Hugging Face Hub or a URL to a deployed Inference Endpoint. Refer to the [Inference Client](https://huggingface.co/docs/huggingface_hub/en/package_reference/inference_client) documentation for more details. 
 
@@ -766,7 +786,7 @@ frames = extractor.extract_frames(text_content=text, entity_key="Diagnosis", str
 
 The ```SentenceFrameExtractor``` instructs the LLM to extract sentence by sentence. The reason is to ensure the accuracy of frame spans. It also prevents LLMs from overseeing sections/ sentences. Empirically, this extractor results in better recall than the ```BasicFrameExtractor``` in complex tasks. 
 
-For concurrent extraction (recommended), the `async/ await` feature is used to speed up inferencing. The `concurrent_batch_size` sets the batch size of sentences to be processed in cocurrent.
+For concurrent extraction (recommended), the `async/await` feature is used to speed up inferencing. The `concurrent_batch_size` sets the batch size of sentences to be processed in cocurrent.
 
 ```python
 from llm_ie.extractors import SentenceFrameExtractor
@@ -775,15 +795,29 @@ extractor = SentenceFrameExtractor(inference_engine, prompt_temp)
 frames = extractor.extract_frames(text_content=text, entity_key="Diagnosis", case_sensitive=False, fuzzy_match=True, concurrent=True, concurrent_batch_size=32)
 ```
 
-The ```multi_turn``` parameter specifies multi-turn conversation for prompting. If True, sentences and LLM outputs will be appended to the input message and carry-over. If False, only the current sentence is prompted. For LLM inference engines that supports prompt cache (e.g., Llama.Cpp, Ollama), use multi-turn conversation prompting can better utilize the KV caching and results in faster inferencing. But for vLLM with [Automatic Prefix Caching (APC)](https://docs.vllm.ai/en/latest/automatic_prefix_caching/apc.html), multi-turn conversation is not necessary.
+The `context_sentences` sets number of sentences before and after the sentence of interest to provide additional context. When `context_sentences=2`, 2 sentences before and 2 sentences after are included in the user prompt as context. When `context_sentences="all"`, the entire document is included as context. When `context_sentences=0`, no context is provided and LLM will only extract based on the current sentence of interest.
 
 ```python
 from llm_ie.extractors import SentenceFrameExtractor
 
-extractor = SentenceFrameExtractor(inference_engine, prompt_temp)
-frames = extractor.extract_frames(text_content=text, entity_key="Diagnosis", multi_turn=False, case_sensitive=False, fuzzy_match=True, stream=True)
+extractor = SentenceFrameExtractor(inference_engine=inference_engine, 
+                                   prompt_template=prompt_temp, 
+                                   context_sentences=2)
+frames = extractor.extract_frames(text_content=text, entity_key="Diagnosis", case_sensitive=False, fuzzy_match=True, stream=True)
 ```
 
+For the sentence:
+
+*The patient has a history of hypertension, hyperlipidemia, and Type 2 diabetes mellitus.*
+
+The context is "previous sentence 2" "previous sentence 1" "the sentence of interest" "proceeding sentence 1" "proceeding sentence 2":
+
+*Emily Brown, MD (Cardiology), Dr. Michael Green, MD (Pulmonology)
+
+*#### Reason for Admission*
+*John Doe, a 49-year-old male, was admitted to the hospital with complaints of chest pain, shortness of breath, and dizziness. The patient has a history of hypertension, hyperlipidemia, and Type 2 diabetes mellitus. #### History of Present Illness*
+*The patient reported that the chest pain started two days prior to admission. The pain was described as a pressure-like sensation in the central chest, radiating to the left arm and jaw.*
+
 </details>
 
 <details>

{llm_ie-0.4.4 → llm_ie-0.4.5}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "llm-ie"
-version = "0.4.4"
+version = "0.4.5"
 description = "An LLM-powered tool that transforms everyday language into robust information extraction pipelines."
 authors = ["Enshuo (David) Hsu"]
 license = "MIT"

{llm_ie-0.4.4 → llm_ie-0.4.5}/src/llm_ie/asset/prompt_guide/SentenceCoTFrameExtractor_prompt_guide.txt

@@ -61,8 +61,8 @@ Example 1 (single entity type with attributes):
     If there is no specific arm, just omit the "Arm" key. If the percentage is not reported, just omit the "Percentage" key. The "Evidence" should always be provided.
 
     # Input placeholder
-    Below is the Adverse reactions section for your reference. I will feed you with sentences from it one by one. 
-    {{input}}
+    Below is the Adverse reactions section:
+    "{{input}}"
 
 
 Example 2 (multiple entity types):
@@ -121,7 +121,7 @@ Example 2 (multiple entity types):
         </Outputs>
 
     # Input placeholder
-    Below is the medical note for your reference. I will feed you with sentences from it one by one.
+    Below is the medical note:
     "{{input}}"
 
 
@@ -213,5 +213,5 @@ Example 3 (multiple entity types with corresponding attributes):
         </Outputs>
 
     # Input placeholder
-    Below is the entire medical note for your reference. I will feed you with sentences from it one by one.
+    Below is the medical note:
     "{{input}}"

{llm_ie-0.4.4 → llm_ie-0.4.5}/src/llm_ie/asset/prompt_guide/SentenceFrameExtractor_prompt_guide.txt

@@ -46,8 +46,8 @@ Example 1 (single entity type with attributes):
     If there is no specific arm, just omit the "Arm" key. If the percentage is not reported, just omit the "Percentage" key. The "Evidence" should always be provided.
 
     # Input placeholder
-    Below is the Adverse reactions section for your reference. I will feed you with sentences from it one by one. 
-    {{input}}
+    Below is the Adverse reactions section:
+    "{{input}}"
 
 
 Example 2 (multiple entity types):
@@ -81,7 +81,7 @@ Example 2 (multiple entity types):
 
 
     # Input placeholder
-    Below is the medical note for your reference. I will feed you with sentences from it one by one.
+    Below is the medical note:
     "{{input}}"
 
 
@@ -141,5 +141,5 @@ Example 3 (multiple entity types with corresponding attributes):
 
 
     # Input placeholder
-    Below is the entire medical note for your reference. I will feed you with sentences from it one by one.
+    Below is the medical note:
     "{{input}}"

{llm_ie-0.4.4 → llm_ie-0.4.5}/src/llm_ie/asset/prompt_guide/SentenceReviewFrameExtractor_prompt_guide.txt

@@ -46,8 +46,8 @@ Example 1 (single entity type with attributes):
     If there is no specific arm, just omit the "Arm" key. If the percentage is not reported, just omit the "Percentage" key. The "Evidence" should always be provided.
 
     # Input placeholder
-    Below is the Adverse reactions section for your reference. I will feed you with sentences from it one by one. 
-    {{input}}
+    Below is the Adverse reactions section:
+    "{{input}}"
 
 
 Example 2 (multiple entity types):
@@ -81,7 +81,7 @@ Example 2 (multiple entity types):
 
 
     # Input placeholder
-    Below is the medical note for your reference. I will feed you with sentences from it one by one.
+    Below is the medical note:
     "{{input}}"
 
 
@@ -141,5 +141,5 @@ Example 3 (multiple entity types with corresponding attributes):
 
 
     # Input placeholder
-    Below is the entire medical note for your reference. I will feed you with sentences from it one by one.
+    Below is the medical note:
     "{{input}}"

{llm_ie-0.4.4 → llm_ie-0.4.5}/src/llm_ie/engines.py

@@ -1,4 +1,5 @@
 import abc
+import warnings
 import importlib
 from typing import List, Dict, Union
 
@@ -242,7 +243,7 @@ class HuggingFaceHubInferenceEngine(InferenceEngine):
         
 
 class OpenAIInferenceEngine(InferenceEngine):
-    def __init__(self, model:str, **kwrs):
+    def __init__(self, model:str, reasoning_model:bool=False, **kwrs):
         """
         The OpenAI API inference engine. Supports OpenAI models and OpenAI compatible servers:
         - vLLM OpenAI compatible server (https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html)
@@ -254,6 +255,8 @@ class OpenAIInferenceEngine(InferenceEngine):
         ----------
         model_name : str
             model name as described in https://platform.openai.com/docs/models
+        reasoning_model : bool, Optional
+            indicator for OpenAI reasoning models ("o" series).
         """
         if importlib.util.find_spec("openai") is None:
             raise ImportError("OpenAI Python API library not found. Please install OpanAI (```pip install openai```).")
@@ -262,6 +265,7 @@ class OpenAIInferenceEngine(InferenceEngine):
         self.client = OpenAI(**kwrs)
         self.async_client = AsyncOpenAI(**kwrs)
         self.model = model
+        self.reasoning_model = reasoning_model
 
     def chat(self, messages:List[Dict[str,str]], max_new_tokens:int=2048, temperature:float=0.0, stream:bool=False, **kwrs) -> str:
         """
@@ -278,14 +282,27 @@ class OpenAIInferenceEngine(InferenceEngine):
         stream : bool, Optional
             if True, LLM generated text will be printed in terminal in real-time. 
         """
-        response = self.client.chat.completions.create(
-            model=self.model,
-            messages=messages,
-            max_tokens=max_new_tokens,
-            temperature=temperature,
-            stream=stream,
-            **kwrs
-        )
+        if self.reasoning_model:
+            if temperature != 0.0:
+                warnings.warn("Reasoning models do not support temperature parameter. Will be ignored.", UserWarning)
+            
+            response = self.client.chat.completions.create(
+                model=self.model,
+                messages=messages,
+                max_completion_tokens=max_new_tokens,
+                stream=stream,
+                **kwrs
+            )
+
+        else:
+            response = self.client.chat.completions.create(
+                model=self.model,
+                messages=messages,
+                max_tokens=max_new_tokens,
+                temperature=temperature,
+                stream=stream,
+                **kwrs
+            )
 
         if stream:
             res = ''
@@ -294,8 +311,17 @@ class OpenAIInferenceEngine(InferenceEngine):
                     if chunk.choices[0].delta.content is not None:
                         res += chunk.choices[0].delta.content
                         print(chunk.choices[0].delta.content, end="", flush=True)
+                    if chunk.choices[0].finish_reason == "length":
+                        warnings.warn("Model stopped generating due to context length limit.", RuntimeWarning)
+                        if self.reasoning_model:
+                            warnings.warn("max_new_tokens includes reasoning tokens and output tokens.", UserWarning)
             return res
         
+        if response.choices[0].finish_reason == "length":
+            warnings.warn("Model stopped generating due to context length limit.", RuntimeWarning)
+            if self.reasoning_model:
+                warnings.warn("max_new_tokens includes reasoning tokens and output tokens.", UserWarning)
+            
         return response.choices[0].message.content
     
 
@@ -303,20 +329,37 @@ class OpenAIInferenceEngine(InferenceEngine):
         """
         Async version of chat method. Streaming is not supported.
         """
-        response = await self.async_client.chat.completions.create(
-            model=self.model,
-            messages=messages,
-            max_tokens=max_new_tokens,
-            temperature=temperature,
-            stream=False,
-            **kwrs
-        )
+        if self.reasoning_model:
+            if temperature != 0.0:
+                warnings.warn("Reasoning models do not support temperature parameter. Will be ignored.", UserWarning)
+            
+            response = await self.async_client.chat.completions.create(
+                model=self.model,
+                messages=messages,
+                max_completion_tokens=max_new_tokens,
+                stream=False,
+                **kwrs
+            )
+        else:
+            response = await self.async_client.chat.completions.create(
+                model=self.model,
+                messages=messages,
+                max_tokens=max_new_tokens,
+                temperature=temperature,
+                stream=False,
+                **kwrs
+            )
         
+        if response.choices[0].finish_reason == "length":
+            warnings.warn("Model stopped generating due to context length limit.", RuntimeWarning)
+            if self.reasoning_model:
+                warnings.warn("max_new_tokens includes reasoning tokens and output tokens.", UserWarning)
+
         return response.choices[0].message.content
     
 
 class AzureOpenAIInferenceEngine(InferenceEngine):
-    def __init__(self, model:str, api_version:str, **kwrs):
+    def __init__(self, model:str, api_version:str, reasoning_model:bool=False, **kwrs):
         """
         The Azure OpenAI API inference engine.
         For parameters and documentation, refer to 
@@ -329,6 +372,8 @@ class AzureOpenAIInferenceEngine(InferenceEngine):
             model name as described in https://platform.openai.com/docs/models
         api_version : str
             the Azure OpenAI API version
+        reasoning_model : bool, Optional
+            indicator for OpenAI reasoning models ("o" series).
         """
         if importlib.util.find_spec("openai") is None:
             raise ImportError("OpenAI Python API library not found. Please install OpanAI (```pip install openai```).")
@@ -340,6 +385,7 @@ class AzureOpenAIInferenceEngine(InferenceEngine):
                                   **kwrs)
         self.async_client = AsyncAzureOpenAI(api_version=self.api_version, 
                                              **kwrs)
+        self.reasoning_model = reasoning_model
 
     def chat(self, messages:List[Dict[str,str]], max_new_tokens:int=2048, temperature:float=0.0, stream:bool=False, **kwrs) -> str:
         """
@@ -356,14 +402,27 @@ class AzureOpenAIInferenceEngine(InferenceEngine):
         stream : bool, Optional
             if True, LLM generated text will be printed in terminal in real-time. 
         """
-        response = self.client.chat.completions.create(
-            model=self.model,
-            messages=messages,
-            max_tokens=max_new_tokens,
-            temperature=temperature,
-            stream=stream,
-            **kwrs
-        )
+        if self.reasoning_model:
+            if temperature != 0.0:
+                warnings.warn("Reasoning models do not support temperature parameter. Will be ignored.", UserWarning)
+            
+            response = self.client.chat.completions.create(
+                model=self.model,
+                messages=messages,
+                max_completion_tokens=max_new_tokens,
+                stream=stream,
+                **kwrs
+            )
+
+        else:
+            response = self.client.chat.completions.create(
+                model=self.model,
+                messages=messages,
+                max_tokens=max_new_tokens,
+                temperature=temperature,
+                stream=stream,
+                **kwrs
+            )
 
         if stream:
             res = ''
@@ -372,8 +431,17 @@ class AzureOpenAIInferenceEngine(InferenceEngine):
                     if chunk.choices[0].delta.content is not None:
                         res += chunk.choices[0].delta.content
                         print(chunk.choices[0].delta.content, end="", flush=True)
+                    if chunk.choices[0].finish_reason == "length":
+                        warnings.warn("Model stopped generating due to context length limit.", RuntimeWarning)
+                        if self.reasoning_model:
+                            warnings.warn("max_new_tokens includes reasoning tokens and output tokens.", UserWarning)
             return res
         
+        if response.choices[0].finish_reason == "length":
+            warnings.warn("Model stopped generating due to context length limit.", RuntimeWarning)
+            if self.reasoning_model:
+                warnings.warn("max_new_tokens includes reasoning tokens and output tokens.", UserWarning)
+
         return response.choices[0].message.content
     
 
@@ -381,15 +449,32 @@ class AzureOpenAIInferenceEngine(InferenceEngine):
         """
         Async version of chat method. Streaming is not supported.
         """
-        response = await self.async_client.chat.completions.create(
-            model=self.model,
-            messages=messages,
-            max_tokens=max_new_tokens,
-            temperature=temperature,
-            stream=False,
-            **kwrs
-        )
+        if self.reasoning_model:
+            if temperature != 0.0:
+                warnings.warn("Reasoning models do not support temperature parameter. Will be ignored.", UserWarning)
+            
+            response = await self.async_client.chat.completions.create(
+                model=self.model,
+                messages=messages,
+                max_completion_tokens=max_new_tokens,
+                stream=False,
+                **kwrs
+            )
+        else:
+            response = await self.async_client.chat.completions.create(
+                model=self.model,
+                messages=messages,
+                max_tokens=max_new_tokens,
+                temperature=temperature,
+                stream=False,
+                **kwrs
+            )
         
+        if response.choices[0].finish_reason == "length":
+            warnings.warn("Model stopped generating due to context length limit.", RuntimeWarning)
+            if self.reasoning_model:
+                warnings.warn("max_new_tokens includes reasoning tokens and output tokens.", UserWarning)
+
         return response.choices[0].message.content
 
     

{llm_ie-0.4.4 → llm_ie-0.4.5}/src/llm_ie/extractors.py

@@ -1,6 +1,5 @@
 import abc
 import re
-import copy
 import json
 import json_repair
 import inspect
@@ -13,7 +12,6 @@ from typing import Set, List, Dict, Tuple, Union, Callable
 from llm_ie.data_types import LLMInformationExtractionFrame, LLMInformationExtractionDocument
 from llm_ie.engines import InferenceEngine
 from colorama import Fore, Style
-from nltk.tokenize import RegexpTokenizer
 
 
 class Extractor:
@@ -139,6 +137,7 @@ class Extractor:
     
 
 class FrameExtractor(Extractor):
+    from nltk.tokenize import RegexpTokenizer
     def __init__(self, inference_engine:InferenceEngine, prompt_template:str, system_prompt:str=None, **kwrs):
         """
         This is the abstract class for frame extraction.
@@ -157,7 +156,8 @@ class FrameExtractor(Extractor):
                          prompt_template=prompt_template,
                          system_prompt=system_prompt,
                          **kwrs)
-        self.tokenizer = RegexpTokenizer(r'\w+|[^\w\s]')
+        
+        self.tokenizer = self.RegexpTokenizer(r'\w+|[^\w\s]')
 
 
     def _jaccard_score(self, s1:Set[str], s2:Set[str]) -> float:
@@ -569,7 +569,8 @@ class ReviewFrameExtractor(BasicFrameExtractor):
 
 class SentenceFrameExtractor(FrameExtractor):
     from nltk.tokenize.punkt import PunktSentenceTokenizer
-    def __init__(self, inference_engine:InferenceEngine, prompt_template:str, system_prompt:str=None, **kwrs):
+    def __init__(self, inference_engine:InferenceEngine, prompt_template:str, system_prompt:str=None,
+                 context_sentences:Union[str, int]="all", **kwrs):
         """
         This class performs sentence-by-sentence information extraction.
         The process is as follows:
@@ -590,10 +591,26 @@ class SentenceFrameExtractor(FrameExtractor):
             prompt template with "{{<placeholder name>}}" placeholder.
         system_prompt : str, Optional
             system prompt.
+        context_sentences : Union[str, int], Optional
+            number of sentences before and after the given sentence to provide additional context. 
+            if "all", the full text will be provided in the prompt as context. 
+            if 0, no additional context will be provided.
+                This is good for tasks that does not require context beyond the given sentence. 
+            if > 0, the number of sentences before and after the given sentence to provide as context.
+                This is good for tasks that require context beyond the given sentence. 
         """
         super().__init__(inference_engine=inference_engine, prompt_template=prompt_template, 
                          system_prompt=system_prompt, **kwrs)
         
+        if not isinstance(context_sentences, int) and context_sentences != "all":
+            raise ValueError('context_sentences must be an integer (>= 0) or "all".')
+        
+        if isinstance(context_sentences, int) and context_sentences < 0:
+            raise ValueError("context_sentences must be a positive integer.")
+            
+        self.context_sentences =context_sentences
+        
+
     def _get_sentences(self, text:str) -> List[Dict[str,str]]:
         """
         This method sentence tokenize the input text into a list of sentences 
@@ -614,9 +631,24 @@ class SentenceFrameExtractor(FrameExtractor):
                             "end": end})    
         return sentences
     
+
+    def _get_context_sentences(self, text_content, i:int, sentences:List[Dict[str, str]], document_key:str=None) -> str:
+        """
+        This function returns the context sentences for the current sentence of interest (i). 
+        """
+        if self.context_sentences == "all":
+            context = text_content if isinstance(text_content, str) else text_content[document_key]
+        elif self.context_sentences == 0:
+            context = ""
+        else:
+            start = max(0, i - self.context_sentences)
+            end = min(i + 1 + self.context_sentences, len(sentences))   
+            context = " ".join([s['sentence_text'] for s in sentences[start:end]])
+        return context
+    
     
     def extract(self, text_content:Union[str, Dict[str,str]], max_new_tokens:int=512, 
-                document_key:str=None, multi_turn:bool=False, temperature:float=0.0, stream:bool=False, **kwrs) -> List[Dict[str,str]]:
+                document_key:str=None, temperature:float=0.0, stream:bool=False, **kwrs) -> List[Dict[str,str]]:
         """
         This method inputs a text and outputs a list of outputs per sentence. 
 
@@ -631,12 +663,6 @@ class SentenceFrameExtractor(FrameExtractor):
         document_key : str, Optional
             specify the key in text_content where document text is. 
             If text_content is str, this parameter will be ignored.
-        multi_turn : bool, Optional
-            multi-turn conversation prompting. 
-            If True, sentences and LLM outputs will be appended to the input message and carry-over. 
-            If False, only the current sentence is prompted. 
-            For LLM inference engines that supports prompt cache (e.g., Llama.Cpp, Ollama), use multi-turn conversation prompting
-            can better utilize the KV caching. 
         temperature : float, Optional
             the temperature for token sampling.
         stream : bool, Optional
@@ -654,19 +680,32 @@ class SentenceFrameExtractor(FrameExtractor):
             if document_key is None:
                 raise ValueError("document_key must be provided when text_content is dict.")
             sentences = self._get_sentences(text_content[document_key])
-        # construct chat messages
-        messages = []
-        if self.system_prompt:
-            messages.append({'role': 'system', 'content': self.system_prompt})
-
-        messages.append({'role': 'user', 'content': self._get_user_prompt(text_content)})
-        messages.append({'role': 'assistant', 'content': 'Sure, please start with the first sentence.'})
 
         # generate sentence by sentence
-        for sent in sentences:
-            messages.append({'role': 'user', 'content': sent['sentence_text']})
+        for i, sent in enumerate(sentences):
+            # construct chat messages
+            messages = []
+            if self.system_prompt:
+                messages.append({'role': 'system', 'content': self.system_prompt})
+
+            context = self._get_context_sentences(text_content, i, sentences, document_key)
+            
+            if self.context_sentences == 0:
+                # no context, just place sentence of interest
+                messages.append({'role': 'user', 'content': self._get_user_prompt(sent['sentence_text'])})
+            else:
+                # insert context
+                messages.append({'role': 'user', 'content': self._get_user_prompt(context)})
+                # simulate conversation
+                messages.append({'role': 'assistant', 'content': 'Sure, please provide the sentence of interest.'})
+                # place sentence of interest
+                messages.append({'role': 'user', 'content': sent['sentence_text']})
+
             if stream:
-                print(f"\n\n{Fore.GREEN}Sentence: {Style.RESET_ALL}\n{sent['sentence_text']}\n")
+                print(f"\n\n{Fore.GREEN}Sentence {i}:{Style.RESET_ALL}\n{sent['sentence_text']}\n")
+                if isinstance(self.context_sentences, int) and self.context_sentences > 0:
+                    print(f"{Fore.YELLOW}Context:{Style.RESET_ALL}\n{context}\n")
+                
                 print(f"{Fore.BLUE}Extraction:{Style.RESET_ALL}")
 
             gen_text = self.inference_engine.chat(
@@ -676,19 +715,13 @@ class SentenceFrameExtractor(FrameExtractor):
                             stream=stream,
                             **kwrs
                         )
-            
-            if multi_turn:
-                # update chat messages with LLM outputs
-                messages.append({'role': 'assistant', 'content': gen_text})
-            else:
-                # delete sentence so that message is reset
-                del messages[-1]
 
             # add to output
             output.append({'sentence_start': sent['start'],
                             'sentence_end': sent['end'],
                             'sentence_text': sent['sentence_text'],
                             'gen_text': gen_text})
+            
         return output
     
 
@@ -726,21 +759,31 @@ class SentenceFrameExtractor(FrameExtractor):
             if document_key is None:
                 raise ValueError("document_key must be provided when text_content is dict.")
             sentences = self._get_sentences(text_content[document_key])
-        # construct chat messages
-        base_messages = []
-        if self.system_prompt:
-            base_messages.append({'role': 'system', 'content': self.system_prompt})
-
-        base_messages.append({'role': 'user', 'content': self._get_user_prompt(text_content)})
-        base_messages.append({'role': 'assistant', 'content': 'Sure, please start with the first sentence.'})
 
         # generate sentence by sentence
         tasks = []
         for i in range(0, len(sentences), concurrent_batch_size):
             batch = sentences[i:i + concurrent_batch_size]
-            for sent in batch:
-                messages = copy.deepcopy(base_messages)
-                messages.append({'role': 'user', 'content': sent['sentence_text']})
+            for j, sent in enumerate(batch):
+                # construct chat messages
+                messages = []
+                if self.system_prompt:
+                    messages.append({'role': 'system', 'content': self.system_prompt})
+
+                context = self._get_context_sentences(text_content, i + j, sentences, document_key)
+                
+                if self.context_sentences == 0:
+                    # no context, just place sentence of interest
+                    messages.append({'role': 'user', 'content': self._get_user_prompt(sent['sentence_text'])})
+                else:
+                    # insert context
+                    messages.append({'role': 'user', 'content': self._get_user_prompt(context)})
+                    # simulate conversation
+                    messages.append({'role': 'assistant', 'content': 'Sure, please provide the sentence of interest.'})
+                    # place sentence of interest
+                    messages.append({'role': 'user', 'content': sent['sentence_text']})
+                
+                # add to tasks
                 task = asyncio.create_task(
                     self.inference_engine.chat_async(
                                 messages=messages, 
@@ -764,10 +807,10 @@ class SentenceFrameExtractor(FrameExtractor):
     
 
     def extract_frames(self, text_content:Union[str, Dict[str,str]], entity_key:str, max_new_tokens:int=512, 
-                            document_key:str=None, multi_turn:bool=False, temperature:float=0.0, stream:bool=False, 
-                            concurrent:bool=False, concurrent_batch_size:int=32,
-                            case_sensitive:bool=False, fuzzy_match:bool=True, fuzzy_buffer_size:float=0.2, fuzzy_score_cutoff:float=0.8,
-                            **kwrs) -> List[LLMInformationExtractionFrame]:
+                        document_key:str=None, temperature:float=0.0, stream:bool=False, 
+                        concurrent:bool=False, concurrent_batch_size:int=32,
+                        case_sensitive:bool=False, fuzzy_match:bool=True, fuzzy_buffer_size:float=0.2, fuzzy_score_cutoff:float=0.8,
+                        **kwrs) -> List[LLMInformationExtractionFrame]:
         """
         This method inputs a text and outputs a list of LLMInformationExtractionFrame
         It use the extract() method and post-process outputs into frames.
@@ -785,12 +828,6 @@ class SentenceFrameExtractor(FrameExtractor):
         document_key : str, Optional
             specify the key in text_content where document text is. 
             If text_content is str, this parameter will be ignored.
-        multi_turn : bool, Optional
-            multi-turn conversation prompting. 
-            If True, sentences and LLM outputs will be appended to the input message and carry-over. 
-            If False, only the current sentence is prompted. 
-            For LLM inference engines that supports prompt cache (e.g., Llama.Cpp, Ollama), use multi-turn conversation prompting
-            can better utilize the KV caching. 
         temperature : float, Optional
             the temperature for token sampling.
         stream : bool, Optional
@@ -815,8 +852,6 @@ class SentenceFrameExtractor(FrameExtractor):
         if concurrent:
             if stream:
                 warnings.warn("stream=True is not supported in concurrent mode.", RuntimeWarning)
-            if multi_turn:
-                warnings.warn("multi_turn=True is not supported in concurrent mode.", RuntimeWarning)
 
             nest_asyncio.apply() # For Jupyter notebook. Terminal does not need this.
             llm_output_sentences = asyncio.run(self.extract_async(text_content=text_content, 
@@ -830,7 +865,6 @@ class SentenceFrameExtractor(FrameExtractor):
             llm_output_sentences = self.extract(text_content=text_content, 
                                             max_new_tokens=max_new_tokens, 
                                             document_key=document_key,
-                                            multi_turn=multi_turn, 
                                             temperature=temperature, 
                                             stream=stream,
                                             **kwrs)
@@ -866,7 +900,8 @@ class SentenceFrameExtractor(FrameExtractor):
 
 class SentenceReviewFrameExtractor(SentenceFrameExtractor):
     def __init__(self, inference_engine:InferenceEngine, prompt_template:str,  
-                 review_mode:str, review_prompt:str=None, system_prompt:str=None, **kwrs):
+                 review_mode:str, review_prompt:str=None, system_prompt:str=None,
+                 context_sentences:Union[str, int]="all", **kwrs):
         """
         This class adds a review step after the SentenceFrameExtractor.
         For each sentence, the review process asks LLM to review its output and:
@@ -888,9 +923,16 @@ class SentenceReviewFrameExtractor(SentenceFrameExtractor):
             addition mode only ask LLM to add new frames, while revision mode ask LLM to regenerate.
         system_prompt : str, Optional
             system prompt.
+        context_sentences : Union[str, int], Optional
+            number of sentences before and after the given sentence to provide additional context. 
+            if "all", the full text will be provided in the prompt as context. 
+            if 0, no additional context will be provided.
+                This is good for tasks that does not require context beyond the given sentence. 
+            if > 0, the number of sentences before and after the given sentence to provide as context.
+                This is good for tasks that require context beyond the given sentence. 
         """
         super().__init__(inference_engine=inference_engine, prompt_template=prompt_template, 
-                         system_prompt=system_prompt, **kwrs)
+                         system_prompt=system_prompt, context_sentences=context_sentences, **kwrs)
 
         if review_mode not in {"addition", "revision"}: 
             raise ValueError('review_mode must be one of {"addition", "revision"}.')
@@ -908,7 +950,7 @@ class SentenceReviewFrameExtractor(SentenceFrameExtractor):
 
 
     def extract(self, text_content:Union[str, Dict[str,str]], max_new_tokens:int=512, 
-                document_key:str=None, multi_turn:bool=False, temperature:float=0.0, stream:bool=False, **kwrs) -> List[Dict[str,str]]:
+                document_key:str=None, temperature:float=0.0, stream:bool=False, **kwrs) -> List[Dict[str,str]]:
         """
         This method inputs a text and outputs a list of outputs per sentence. 
 
@@ -923,12 +965,6 @@ class SentenceReviewFrameExtractor(SentenceFrameExtractor):
         document_key : str, Optional
             specify the key in text_content where document text is. 
             If text_content is str, this parameter will be ignored.
-        multi_turn : bool, Optional
-            multi-turn conversation prompting. 
-            If True, sentences and LLM outputs will be appended to the input message and carry-over. 
-            If False, only the current sentence is prompted. 
-            For LLM inference engines that supports prompt cache (e.g., Llama.Cpp, Ollama), use multi-turn conversation prompting
-            can better utilize the KV caching. 
         temperature : float, Optional
             the temperature for token sampling.
         stream : bool, Optional
@@ -946,19 +982,31 @@ class SentenceReviewFrameExtractor(SentenceFrameExtractor):
             if document_key is None:
                 raise ValueError("document_key must be provided when text_content is dict.")
             sentences = self._get_sentences(text_content[document_key])
-        # construct chat messages
-        messages = []
-        if self.system_prompt:
-            messages.append({'role': 'system', 'content': self.system_prompt})
+        
+        # generate sentence by sentence
+        for i, sent in enumerate(sentences):
+            # construct chat messages
+            messages = []
+            if self.system_prompt:
+                messages.append({'role': 'system', 'content': self.system_prompt})
 
-        messages.append({'role': 'user', 'content': self._get_user_prompt(text_content)})
-        messages.append({'role': 'assistant', 'content': 'Sure, please start with the first sentence.'})
+            context = self._get_context_sentences(text_content, i, sentences, document_key)
+            
+            if self.context_sentences == 0:
+                # no context, just place sentence of interest
+                messages.append({'role': 'user', 'content': self._get_user_prompt(sent['sentence_text'])})
+            else:
+                # insert context
+                messages.append({'role': 'user', 'content': self._get_user_prompt(context)})
+                # simulate conversation
+                messages.append({'role': 'assistant', 'content': 'Sure, please provide the sentence of interest.'})
+                # place sentence of interest
+                messages.append({'role': 'user', 'content': sent['sentence_text']})
 
-        # generate sentence by sentence
-        for sent in sentences:
-            messages.append({'role': 'user', 'content': sent['sentence_text']})
             if stream:
-                print(f"\n\n{Fore.GREEN}Sentence: {Style.RESET_ALL}\n{sent['sentence_text']}\n")
+                print(f"\n\n{Fore.GREEN}Sentence {i}: {Style.RESET_ALL}\n{sent['sentence_text']}\n")
+                if isinstance(self.context_sentences, int) and self.context_sentences > 0:
+                    print(f"{Fore.YELLOW}Context:{Style.RESET_ALL}\n{context}\n")
                 print(f"{Fore.BLUE}Initial Output:{Style.RESET_ALL}")
 
             initial = self.inference_engine.chat(
@@ -988,13 +1036,6 @@ class SentenceReviewFrameExtractor(SentenceFrameExtractor):
                 gen_text = review
             elif self.review_mode == "addition":
                 gen_text = initial + '\n' + review
-            
-            if multi_turn:
-                # update chat messages with LLM outputs
-                messages.append({'role': 'assistant', 'content': review})
-            else:
-                # delete sentence and review so that message is reset
-                del messages[-3:]
 
             # add to output
             output.append({'sentence_start': sent['start'],
@@ -1040,24 +1081,33 @@ class SentenceReviewFrameExtractor(SentenceFrameExtractor):
             if document_key is None:
                 raise ValueError("document_key must be provided when text_content is dict.")
             sentences = self._get_sentences(text_content[document_key])
-        # construct chat messages
-        base_messages = []
-        if self.system_prompt:
-            base_messages.append({'role': 'system', 'content': self.system_prompt})
-
-        base_messages.append({'role': 'user', 'content': self._get_user_prompt(text_content)})
-        base_messages.append({'role': 'assistant', 'content': 'Sure, please start with the first sentence.'})
 
         # generate initial outputs sentence by sentence
-        initials = []
         tasks = []
-        message_list = []
+        messages_list = []
         for i in range(0, len(sentences), concurrent_batch_size):
             batch = sentences[i:i + concurrent_batch_size]
-            for sent in batch:
-                messages = copy.deepcopy(base_messages)
-                messages.append({'role': 'user', 'content': sent['sentence_text']})
-                message_list.append(messages)
+            for j, sent in enumerate(batch):
+                # construct chat messages
+                messages = []
+                if self.system_prompt:
+                    messages.append({'role': 'system', 'content': self.system_prompt})
+
+                context = self._get_context_sentences(text_content, i + j, sentences, document_key)
+                
+                if self.context_sentences == 0:
+                    # no context, just place sentence of interest
+                    messages.append({'role': 'user', 'content': self._get_user_prompt(sent['sentence_text'])})
+                else:
+                    # insert context
+                    messages.append({'role': 'user', 'content': self._get_user_prompt(context)})
+                    # simulate conversation
+                    messages.append({'role': 'assistant', 'content': 'Sure, please provide the sentence of interest.'})
+                    # place sentence of interest
+                    messages.append({'role': 'user', 'content': sent['sentence_text']})
+
+                messages_list.append(messages)
+
                 task = asyncio.create_task(
                     self.inference_engine.chat_async(
                                 messages=messages, 
@@ -1071,15 +1121,15 @@ class SentenceReviewFrameExtractor(SentenceFrameExtractor):
         # Wait until the batch is done, collect results and move on to next batch
         responses = await asyncio.gather(*tasks)
         # Collect initials
-        for gen_text, sent, message in zip(responses, sentences, message_list):
+        initials = []
+        for gen_text, sent, messages in zip(responses, sentences, messages_list):
             initials.append({'sentence_start': sent['start'],
                             'sentence_end': sent['end'],
                             'sentence_text': sent['sentence_text'],
                             'gen_text': gen_text,
-                            'messages': message})
-            
+                            'messages': messages})
+
         # Review
-        reviews = []
         tasks = []
         for i in range(0, len(initials), concurrent_batch_size):
             batch = initials[i:i + concurrent_batch_size]
@@ -1101,6 +1151,7 @@ class SentenceReviewFrameExtractor(SentenceFrameExtractor):
             responses = await asyncio.gather(*tasks)
 
         # Collect reviews
+        reviews = []
         for gen_text, sent in zip(responses, sentences):
             reviews.append({'sentence_start': sent['start'],
                             'sentence_end': sent['end'],
@@ -1123,7 +1174,8 @@ class SentenceReviewFrameExtractor(SentenceFrameExtractor):
 
 class SentenceCoTFrameExtractor(SentenceFrameExtractor):
     from nltk.tokenize.punkt import PunktSentenceTokenizer
-    def __init__(self, inference_engine:InferenceEngine, prompt_template:str, system_prompt:str=None, **kwrs):
+    def __init__(self, inference_engine:InferenceEngine, prompt_template:str, system_prompt:str=None, 
+                 context_sentences:Union[str, int]="all", **kwrs):
         """
         This class performs sentence-based Chain-of-thoughts (CoT) information extraction.
         A simulated chat follows this process:
@@ -1145,13 +1197,20 @@ class SentenceCoTFrameExtractor(SentenceFrameExtractor):
             prompt template with "{{<placeholder name>}}" placeholder.
         system_prompt : str, Optional
             system prompt.
+        context_sentences : Union[str, int], Optional
+            number of sentences before and after the given sentence to provide additional context. 
+            if "all", the full text will be provided in the prompt as context. 
+            if 0, no additional context will be provided.
+                This is good for tasks that does not require context beyond the given sentence. 
+            if > 0, the number of sentences before and after the given sentence to provide as context.
+                This is good for tasks that require context beyond the given sentence. 
         """
         super().__init__(inference_engine=inference_engine, prompt_template=prompt_template, 
-                         system_prompt=system_prompt, **kwrs)
+                         system_prompt=system_prompt, context_sentences=context_sentences, **kwrs)
         
 
     def extract(self, text_content:Union[str, Dict[str,str]], max_new_tokens:int=512, 
-                document_key:str=None, multi_turn:bool=False, temperature:float=0.0, stream:bool=False, **kwrs) -> List[Dict[str,str]]:
+                document_key:str=None, temperature:float=0.0, stream:bool=False, **kwrs) -> List[Dict[str,str]]:
         """
         This method inputs a text and outputs a list of outputs per sentence. 
 
@@ -1166,12 +1225,6 @@ class SentenceCoTFrameExtractor(SentenceFrameExtractor):
         document_key : str, Optional
             specify the key in text_content where document text is. 
             If text_content is str, this parameter will be ignored.
-        multi_turn : bool, Optional
-            multi-turn conversation prompting. 
-            If True, sentences and LLM outputs will be appended to the input message and carry-over. 
-            If False, only the current sentence is prompted. 
-            For LLM inference engines that supports prompt cache (e.g., Llama.Cpp, Ollama), use multi-turn conversation prompting
-            can better utilize the KV caching. 
         temperature : float, Optional
             the temperature for token sampling.
         stream : bool, Optional
@@ -1187,19 +1240,31 @@ class SentenceCoTFrameExtractor(SentenceFrameExtractor):
             sentences = self._get_sentences(text_content)
         elif isinstance(text_content, dict):
             sentences = self._get_sentences(text_content[document_key])
-        # construct chat messages
-        messages = []
-        if self.system_prompt:
-            messages.append({'role': 'system', 'content': self.system_prompt})
-
-        messages.append({'role': 'user', 'content': self._get_user_prompt(text_content)})
-        messages.append({'role': 'assistant', 'content': 'Sure, please start with the first sentence.'})
 
         # generate sentence by sentence
-        for sent in sentences:
-            messages.append({'role': 'user', 'content': sent['sentence_text']})
+        for i, sent in enumerate(sentences):
+            # construct chat messages
+            messages = []
+            if self.system_prompt:
+                messages.append({'role': 'system', 'content': self.system_prompt})
+
+            context = self._get_context_sentences(text_content, i, sentences, document_key)
+            
+            if self.context_sentences == 0:
+                # no context, just place sentence of interest
+                messages.append({'role': 'user', 'content': self._get_user_prompt(sent['sentence_text'])})
+            else:
+                # insert context
+                messages.append({'role': 'user', 'content': self._get_user_prompt(context)})
+                # simulate conversation
+                messages.append({'role': 'assistant', 'content': 'Sure, please provide the sentence of interest.'})
+                # place sentence of interest
+                messages.append({'role': 'user', 'content': sent['sentence_text']})
+
             if stream:
                 print(f"\n\n{Fore.GREEN}Sentence: {Style.RESET_ALL}\n{sent['sentence_text']}\n")
+                if isinstance(self.context_sentences, int) and self.context_sentences > 0:
+                    print(f"{Fore.YELLOW}Context:{Style.RESET_ALL}\n{context}\n")
                 print(f"{Fore.BLUE}CoT:{Style.RESET_ALL}")
 
             gen_text = self.inference_engine.chat(
@@ -1209,13 +1274,6 @@ class SentenceCoTFrameExtractor(SentenceFrameExtractor):
                             stream=stream,
                             **kwrs
                         )
-            
-            if multi_turn:
-                # update chat messages with LLM outputs
-                messages.append({'role': 'assistant', 'content': gen_text})
-            else:
-                # delete sentence so that message is reset
-                del messages[-1]
 
             # add to output
             output.append({'sentence_start': sent['start'],