fix(ParseNode): leave room for LLM reply in context window

Author: f-aguzzi

pyproject.toml

@@ -14,12 +14,12 @@ authors = [
 ]
 
 dependencies = [
-    "langchain==0.2.14",
+    "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==0.1.22",
+    "langchain-openai>=0.1.22",
     "langchain-groq>=0.1.3",
     "langchain-aws>=0.1.3",
     "langchain-anthropic>=0.1.11",

requirements-dev.lock

@@ -179,6 +179,7 @@ graphviz==0.20.3
     # via scrapegraphai
 greenlet==3.0.3
     # via playwright
+    # via sqlalchemy
 groq==0.9.0
     # via langchain-groq
 grpc-google-iam-v1==0.13.1

requirements.lock

@@ -133,6 +133,7 @@ graphviz==0.20.3
     # via scrapegraphai
 greenlet==3.0.3
     # via playwright
+    # via sqlalchemy
 groq==0.9.0
     # via langchain-groq
 grpc-google-iam-v1==0.13.1

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/nodes/parse_node.py

@@ -6,7 +6,6 @@
 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):
@@ -79,16 +78,18 @@ 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]