Merge pull request #562 from ScrapeGraphAI/support_structured_output_shema_openai

Support structured output shema openai

Author: VinciGit00

examples/anthropic/search_graph_schema_haiku.py

@@ -3,13 +3,12 @@
 """
 
 import os
+from typing import List
 from dotenv import load_dotenv
-load_dotenv()
-
+from pydantic import BaseModel, Field
 from scrapegraphai.graphs import SearchGraph
 
-from pydantic import BaseModel, Field
-from typing import List
+load_dotenv()
 
 # ************************************************
 # Define the output schema for the graph

examples/azure/smart_scraper_schema_azure.py

@@ -2,7 +2,8 @@
 Basic example of scraping pipeline using SmartScraper with schema
 """
 
-import os, json
+import os
+import json
 from typing import List
 from pydantic import BaseModel, Field
 from dotenv import load_dotenv

examples/local_models/smart_scraper_schema_ollama.py

@@ -19,7 +19,7 @@ class Projects(BaseModel):
 
 graph_config = {
     "llm": {
-        "model": "ollama/llama3",
+        "model": "ollama/llama3.1",
         "temperature": 0,
         "format": "json",  # Ollama needs the format to be specified explicitly
         # "base_url": "http://localhost:11434", # set ollama URL arbitrarily

examples/openai/smart_scraper_openai.py

@@ -2,7 +2,8 @@
 Basic example of scraping pipeline using SmartScraper
 """
 
-import os, json
+import os
+import json
 from scrapegraphai.graphs import SmartScraperGraph
 from scrapegraphai.utils import prettify_exec_info
 from dotenv import load_dotenv

examples/openai/smart_scraper_schema_openai.py

@@ -30,7 +30,7 @@ class Projects(BaseModel):
 graph_config = {
     "llm": {
         "api_key":openai_key,
-        "model": "gpt-4o",
+        "model": "gpt-4o-mini",
     },
     "verbose": True,
     "headless": False,

pyproject.toml

@@ -14,12 +14,12 @@ authors = [
 ]
 
 dependencies = [
-    "langchain>=0.2.10",
+    "langchain>=0.2.14",
     "langchain-fireworks>=0.1.3",
     "langchain_community>=0.2.9",
     "langchain-google-genai>=1.0.7",
     "langchain-google-vertexai>=1.0.7",
-    "langchain-openai>=0.1.17",
+    "langchain-openai>=0.1.22",
     "langchain-groq>=0.1.3",
     "langchain-aws>=0.1.3",
     "langchain-anthropic>=0.1.11",

requirements-dev.lock

@@ -255,7 +255,7 @@ jsonschema-specifications==2023.12.1
     # via jsonschema
 kiwisolver==1.4.5
     # via matplotlib
-langchain==0.2.12
+langchain==0.2.14
     # via langchain-community
     # via scrapegraphai
 langchain-anthropic==0.1.22
@@ -264,7 +264,7 @@ langchain-aws==0.1.16
     # via scrapegraphai
 langchain-community==0.2.11
     # via scrapegraphai
-langchain-core==0.2.29
+langchain-core==0.2.33
     # via langchain
     # via langchain-anthropic
     # via langchain-aws
@@ -292,7 +292,7 @@ langchain-mistralai==0.1.12
     # via scrapegraphai
 langchain-nvidia-ai-endpoints==0.2.1
     # via scrapegraphai
-langchain-openai==0.1.21
+langchain-openai==0.1.22
     # via scrapegraphai
 langchain-text-splitters==0.2.2
     # via langchain

requirements.lock

@@ -178,6 +178,7 @@ jinja2==3.1.4
     # via torch
 jiter==0.5.0
     # via anthropic
+    # via openai
 jmespath==1.0.1
     # via boto3
     # via botocore
@@ -187,7 +188,7 @@ jsonpatch==1.33
     # via langchain-core
 jsonpointer==3.0.0
     # via jsonpatch
-langchain==0.2.11
+langchain==0.2.14
     # via langchain-community
     # via scrapegraphai
 langchain-anthropic==0.1.20
@@ -196,7 +197,7 @@ langchain-aws==0.1.12
     # via scrapegraphai
 langchain-community==0.2.10
     # via scrapegraphai
-langchain-core==0.2.28
+langchain-core==0.2.33
     # via langchain
     # via langchain-anthropic
     # via langchain-aws
@@ -224,7 +225,7 @@ langchain-mistralai==0.1.12
     # via scrapegraphai
 langchain-nvidia-ai-endpoints==0.1.7
     # via scrapegraphai
-langchain-openai==0.1.17
+langchain-openai==0.1.22
     # via scrapegraphai
 langchain-text-splitters==0.2.2
     # via langchain
@@ -264,7 +265,7 @@ numpy==1.26.4
     # via sentence-transformers
     # via shapely
     # via transformers
-openai==1.37.0
+openai==1.41.0
     # via langchain-fireworks
     # via langchain-openai
 orjson==3.10.6

requirements.txt

@@ -1,9 +1,9 @@
-langchain>=0.2.10
+langchain>=0.2.14
 langchain-fireworks>=0.1.3
 langchain_community>=0.2.9
 langchain-google-genai>=1.0.7
 langchain-google-vertexai>=1.0.7
-langchain-openai>=0.1.17
+langchain-openai>=0.1.22
 langchain-groq>=0.1.3
 langchain-aws>=0.1.3
 langchain-anthropic>=0.1.11

scrapegraphai/helpers/models_tokens.py

@@ -43,7 +43,7 @@
         "gpt-4-32k-0613": 32768,
         "gpt-4o": 128000,
         "gpt-4o-mini":128000,
-        "chatgpt-4o-latest":128000
+        "chatgpt-4o-latest": 128000
     },
     "google_genai": {
         "gemini-pro": 128000,

scrapegraphai/nodes/generate_answer_node.py

@@ -5,7 +5,12 @@
 from langchain.prompts import PromptTemplate
 from langchain_core.output_parsers import JsonOutputParser
 from langchain_core.runnables import RunnableParallel
-from langchain_openai import ChatOpenAI
+from langchain_openai import ChatOpenAI, AzureChatOpenAI
+from langchain_mistralai import ChatMistralAI
+from langchain_anthropic import ChatAnthropic
+from langchain_groq import ChatGroq
+from langchain_fireworks import ChatFireworks
+from langchain_google_vertexai import ChatVertexAI
 from langchain_community.chat_models import ChatOllama
 from tqdm import tqdm
 from ..utils.logging import get_logger
@@ -88,12 +93,19 @@ def execute(self, state: dict) -> dict:
         # Initialize the output parser
         if self.node_config.get("schema", None) is not None:
             output_parser = JsonOutputParser(pydantic_object=self.node_config["schema"])
+
+            # Use built-in structured output for providers that allow it
+            if isinstance(self.llm_model, (ChatOpenAI, ChatMistralAI, ChatAnthropic, ChatFireworks, ChatGroq, ChatVertexAI)):
+                self.llm_model = self.llm_model.with_structured_output(
+                    schema = self.node_config["schema"],
+                    method="json_schema")
+
         else:
             output_parser = JsonOutputParser()
 
         format_instructions = output_parser.get_format_instructions()
 
-        if  isinstance(self.llm_model, ChatOpenAI) and not self.script_creator or self.force and not self.script_creator or self.is_md_scraper:
+        if isinstance(self.llm_model, (ChatOpenAI, AzureChatOpenAI)) and not self.script_creator or self.force and not self.script_creator or self.is_md_scraper:
             template_no_chunks_prompt  = TEMPLATE_NO_CHUNKS_MD
             template_chunks_prompt  = TEMPLATE_CHUNKS_MD
             template_merge_prompt  = TEMPLATE_MERGE_MD

scrapegraphai/nodes/parse_node.py

@@ -5,7 +5,6 @@
 from semchunk import chunk
 from langchain_community.document_transformers import Html2TextTransformer
 from langchain_core.documents import Document
-from ..utils.logging import get_logger
 from .base_node import BaseNode
 
 class ParseNode(BaseNode):
@@ -78,16 +77,18 @@ def execute(self, state: dict) -> dict:
         else:
             docs_transformed = docs_transformed[0]
 
+            # Adapt the chunk size, leaving room for the reply, the prompt and the schema
+            chunk_size = self.node_config.get("chunk_size", 4096)
+            chunk_size = min(chunk_size - 500, int(chunk_size * 0.9))
+
             if isinstance(docs_transformed, Document):
-                
                 chunks = chunk(text=docs_transformed.page_content,
-                            chunk_size=self.node_config.get("chunk_size", 4096)-250,
+                            chunk_size=chunk_size,
                             token_counter=lambda text: len(text.split()),
                             memoize=False)
             else:
-
                 chunks = chunk(text=docs_transformed,
-                                chunk_size=self.node_config.get("chunk_size", 4096)-250,
+                                chunk_size=chunk_size,
                                 token_counter=lambda text: len(text.split()),
                                 memoize=False)
 

scrapegraphai/utils/token_calculator.py

@@ -1,5 +1,5 @@
 """ 
-Module for truncatinh in chunks the messages
+Module for truncating in chunks the messages
 """
 from typing import List
 import tiktoken
@@ -27,7 +27,7 @@ def truncate_text_tokens(text: str, model: str, encoding_name: str) -> List[str]
     """
 
     encoding = tiktoken.get_encoding(encoding_name)
-    max_tokens = models_tokens[model] - 500
+    max_tokens = min(models_tokens[model] - 500, int(models_tokens[model] * 0.9))
     encoded_text = encoding.encode(text)
 
     chunks = [encoded_text[i:i + max_tokens]