LLM-Driven Extraction of Unstructured Data — Built for API Deployments & ETL Pipeline Workflows
Python ai-agentsdata-engineeringdocument-aigenerative-aiidpjson-extractionllmmcp-serverocrpdf-extractionprompt-engineeringstructured-output
updated classifier tool with latest sdk version (#152)
alid8f5124753cc
diff --git a/backend/tool_instance/tool_instance_helper.py b/backend/tool_instance/tool_instance_helper.pyindex 07b59b03c..fd3ce63b5 100644--- a/backend/tool_instance/tool_instance_helper.py+++ b/backend/tool_instance/tool_instance_helper.py@@ -367,7 +367,7 @@ class ToolInstanceHelper:user: User, tool_uid: str, tool_meta: dict[str, Any]) -> None:tool: Tool = ToolProcessor.get_tool_by_uid(tool_uid=tool_uid)- adapter_ids = set()+ adapter_ids: set[str] = set()for llm in tool.properties.adapter.language_models:if llm.is_enabled and llm.adapter_id:@@ -413,7 +413,7 @@ class ToolInstanceHelper:@staticmethoddef validate_adapter_access(user: User,- adapter_ids: str,+ adapter_ids: list[str],) -> None:adapter_instances = AdapterInstance.objects.filter(id__in=adapter_idsdiff --git a/tools/classifier/Dockerfile b/tools/classifier/Dockerfileindex 4c95f3d4d..a8b1f90a5 100644--- a/tools/classifier/Dockerfile+++ b/tools/classifier/Dockerfile@@ -4,7 +4,7 @@ LABEL maintainer="Zipstack Inc."ENV UNSTRACT_ENTRYPOINT "python /app/src/main.py"# Install dependencies for unstructured library's partition-RUN apt-get update && apt-get --no-install-recommends -y install libmagic-dev poppler-utils tesseract-ocr libreoffice pandoc\+RUN apt-get update && apt-get --no-install-recommends -y install libmagic-dev poppler-utils\&& apt-get clean \&& rm -rf /var/lib/apt/lists/*diff --git a/tools/classifier/requirements.txt b/tools/classifier/requirements.txtindex 2aa4cfe0d..7ecf41515 100644--- a/tools/classifier/requirements.txt+++ b/tools/classifier/requirements.txt@@ -1,6 +1,4 @@# Add your dependencies here-filetype==1.2.0-unstructured[all-docs]==0.10.10# Required for all unstract tools-unstract-sdk~=0.10.0+unstract-sdk~=0.15.2diff --git a/tools/classifier/src/config/properties.json b/tools/classifier/src/config/properties.jsonindex 38b425fef..b1139b9ae 100644--- a/tools/classifier/src/config/properties.json+++ b/tools/classifier/src/config/properties.json@@ -2,7 +2,7 @@"schemaVersion": "0.0.1","displayName": "File Classifier","functionName": "classify",- "toolVersion": "0.0.1",+ "toolVersion": "0.0.2","description": "Classifies a file into a bin based on its contents","input": {"description": "File to be classified"@@ -33,6 +33,15 @@{"isEnabled": false}+ ],+ "textExtractors": [+ {+ "isEnabled": true,+ "adapterId": "textExtractorId",+ "title": "Text Extraction Adapter",+ "isRequired": false,+ "description": "Choose extractor adapter to extract text"+ }]},"ioCompatibility": {diff --git a/tools/classifier/src/helper.py b/tools/classifier/src/helper.pyindex 6f31d881b..c75ede6f5 100644--- a/tools/classifier/src/helper.py+++ b/tools/classifier/src/helper.py@@ -1,7 +1,11 @@-import io+from typing import Any, Optional+from llama_index.llms import LLM+from unstract.sdk.cache import ToolCache+from unstract.sdk.constants import ToolEnvfrom unstract.sdk.tool.base import BaseTool-from unstructured.partition.auto import partition+from unstract.sdk.utils import ToolUtils+from unstract.sdk.x2txt import X2Textclass ClassifierHelper:@@ -10,7 +14,9 @@ class ClassifierHelper:def __init__(self, tool: BaseTool) -> None:self.tool = tool- def extract_text(self, file: str) -> str:+ def extract_text(+ self, file: str, text_extraction_adapter_id: Optional[str]+ ) -> Optional[str]:"""Extract text from file.Args:@@ -19,11 +25,159 @@ class ClassifierHelper:Returns:str: page content"""+ if not text_extraction_adapter_id:+ return self._extract_from_file(file)++ return self._extract_from_adapter(file, text_extraction_adapter_id)++ def _extract_from_adapter(+ self, file: str, adapter_id: str+ ) -> Optional[str]:+ """Extract text from adapter.++ Args:+ file: The path to the input file+ adapter_id: The id of the adapter+ Returns:+ str: page content+ """+ self.tool.stream_log(+ "Creating text extraction adapter "+ f"using adapter_id: {adapter_id}"+ )+ tool_extraction = X2Text(tool=self.tool)+ adapter = tool_extraction.get_x2text(adapter_id)++ self.tool.stream_log(+ "Text extraction adapter has been created successfully."+ )+ self.tool.stream_log("Adapter created")++ try:+ extracted_text: str = adapter.process(input_file_path=file)+ return extracted_text+ except Exception as e:+ self.tool.stream_log(f"Adapter error: {e}")+ return None++ def _extract_from_file(self, file: str) -> Optional[str]:+ """Extract text from file.++ Args:+ file: The path to the input file+ Returns:+ str: page content+ """+ self.tool.stream_log("Extracting text from file")try:- with open(file, mode="rb") as input_file_obj:- bytes_io = io.BytesIO(input_file_obj.read())- elements = partition(file=bytes_io)+ with open(file, "rb") as f:+ text = f.read().decode("utf-8")except Exception as e:- self.tool.stream_error_and_exit(f"Error partitioning file: {e}")- text = "\n\n".join([str(el) for el in elements])+ self.tool.stream_log(f"File error: {e}")+ return None++ self.tool.stream_log("Text extracted from file")return text++ def find_classification(+ self,+ use_cache: bool,+ settings_string: str,+ bins: list[str],+ prompt: str,+ llm: LLM,+ ) -> Optional[str]:+ """Find classification for text.+ Args:+ use_cache (bool): Whether to use cache+ settings_string (str): hash of settings+ prompt (str): Prompt+ bins (list[str]): Classification Bins+ llm (LLM): LLM++ Returns:+ Optional[str]: Classification or None if not found in cache.+ If found in cache, returns the classification from the cache.+ If not found in cache, calls the LLM and returns the+ classification.+ """+ classification = None+ cache_key = None+ if use_cache:+ cache_key = (+ f"cache:{self.tool.workflow_id}:"+ f"{ToolUtils.hash_str(settings_string)}:"+ f"{ToolUtils.hash_str(prompt)}"+ )+ classification = self.get_result_from_cache(cache_key=cache_key)++ if classification is None:+ self.tool.stream_log(+ "No classification found in cache, calling LLM."+ )+ classification = self.call_llm(prompt=prompt, llm=llm)+ if not classification:+ classification = "unknown"+ classification = classification.strip().lower()+ bins = [bin.lower() for bin in bins]+ if classification not in bins:+ self.tool.stream_error_and_exit(+ f"Invalid classification done: {classification}"+ )+ return classification++ def call_llm(self, prompt: str, llm: LLM) -> Optional[str]:+ """Call LLM.++ Args:+ prompt (str): Prompt+ llm (LLM): LLM++ Returns:+ str: Classification+ """+ try:+ response = llm.complete(prompt, max_tokens=50, stop=["\n"])+ classification: str = response.text.strip()+ self.tool.stream_log(f"LLM response: {response}")+ return classification+ except Exception as e:+ self.tool.stream_error_and_exit(f"Error calling LLM {e}")+ return None++ def get_result_from_cache(self, cache_key: str) -> Optional[str]:+ """Get result from cache.++ Args:+ cache_key (str): key++ Returns:+ Optional[str]: result+ """+ cache = ToolCache(+ tool=self.tool,+ platform_host=self.tool.get_env_or_die(ToolEnv.PLATFORM_HOST),+ platform_port=int(self.tool.get_env_or_die(ToolEnv.PLATFORM_PORT)),+ )+ cached_response: Optional[str] = cache.get(cache_key)+ if cached_response is not None:+ classification = cached_response+ self.tool.stream_cost(cost=0.0, cost_units="cache")+ return classification+ return None++ def save_result_to_cache(self, cache_key: str, result: Any) -> None:+ """Save result to cache.++ Args:+ cache_key (str): key+ result (Any): result to save in cache+ Returns:+ None: None+ """+ cache = ToolCache(+ tool=self.tool,+ platform_host=self.tool.get_env_or_die(ToolEnv.PLATFORM_HOST),+ platform_port=int(self.tool.get_env_or_die(ToolEnv.PLATFORM_PORT)),+ )+ cache.set(cache_key, result)diff --git a/tools/classifier/src/main.py b/tools/classifier/src/main.pyindex 2eafa0d8c..77d482cef 100644--- a/tools/classifier/src/main.py+++ b/tools/classifier/src/main.py@@ -1,18 +1,18 @@import shutilimport sysfrom pathlib import Path-from typing import Any--from llama_index import set_global_service_context-from unstract.sdk.cache import ToolCache-from unstract.sdk.constants import LogLevel, LogState, MetadataKey, ToolEnv+from typing import Any, Optional++from helper import ClassifierHelper # type: ignore+from unstract.sdk.constants import (+ LogLevel,+ LogState,+ MetadataKey,+ ToolSettingsKey,+)from unstract.sdk.llm import ToolLLMfrom unstract.sdk.tool.base import BaseToolfrom unstract.sdk.tool.entrypoint import ToolEntrypoint-from unstract.sdk.utils import ToolUtils-from unstract.sdk.utils.service_context import ServiceContext--from .helper import ClassifierHelperclass UnstractClassifier(BaseTool):@@ -21,10 +21,14 @@ class UnstractClassifier(BaseTool):self.helper = ClassifierHelper(tool=self)def validate(self, input_file: str, settings: dict[str, Any]) -> None:- bins = settings["classificationBins"]-- if len(bins) < 2:+ bins: Optional[list[str]] = settings.get("classificationBins")+ llm_adapter_instance_id = settings.get(ToolSettingsKey.LLM_ADAPTER_ID)+ if not bins:+ self.stream_error_and_exit("Classification bins are required")+ elif len(bins) < 2:self.stream_error_and_exit("At least two bins are required")+ if not llm_adapter_instance_id:+ self.stream_error_and_exit("Choose a llm to process the classifier")def run(self,@@ -34,6 +38,8 @@ class UnstractClassifier(BaseTool):) -> None:bins = settings["classificationBins"]use_cache = settings["useCache"]+ text_extraction_adapter_id = settings["textExtractorId"]+ llm_adapter_instance_id = settings[ToolSettingsKey.LLM_ADAPTER_ID]# Update GUIinput_log = f"### Classification bins:\n```text\n{bins}\n```\n\n"@@ -41,14 +47,20 @@ class UnstractClassifier(BaseTool):self.stream_update(input_log, state=LogState.INPUT_UPDATE)self.stream_update(output_log, state=LogState.OUTPUT_UPDATE)- self.stream_log("Reading file...")- text = self.helper.extract_text(input_file)+ self.stream_log(f"Reading file... {input_file}")+ text: Optional[str] = self.helper.extract_text(+ file=input_file,+ text_extraction_adapter_id=text_extraction_adapter_id,+ )+ if not text:+ self.stream_error_and_exit("Unable to extract text")+ returnself.stream_log(f"Text length: {len(text)}")# Update GUIinput_text_for_log = textif len(input_text_for_log) > 500:- input_text_for_log = input_text_for_log[:500] + "...(truncated)"+ input_text_for_log = input_text_for_log[:50000] + "...(truncated)"input_log = (f"### Classification bins:\n```text\n{bins}\n```\n\n"f"### Input text:\n\n```text\n{input_text_for_log}\n```\n\n"@@ -61,17 +73,10 @@ class UnstractClassifier(BaseTool):bins.append("unknown")bins_with_quotes = [f"'{b}'" for b in bins]- tool_llm = ToolLLM(tool=self, tool_settings=settings)- llm = tool_llm.get_llm()+ tool_llm = ToolLLM(tool=self)+ llm = tool_llm.get_llm(adapter_instance_id=llm_adapter_instance_id)if not llm:self.stream_error_and_exit("Unable to get llm instance")- service_context = ServiceContext.get_service_context(- platform_api_key=self.get_env_or_die(ToolEnv.PLATFORM_API_KEY),- llm=llm,- workflow_id=self.workflow_id,- execution_id=self.execution_id,- )- set_global_service_context(service_context)max_tokens = tool_llm.get_max_tokens(reserved_for_output=50 + 1000)max_bytes = int(max_tokens * 1.3)@@ -93,60 +98,19 @@ class UnstractClassifier(BaseTool):prompt = (f"Classify the following text into one of the following categories: {' '.join(bins_with_quotes)}.\n\n" # noqa: E501f"Your categorization should be strictly exactly one of the items in the " # noqa: E501- f"Your categorization should be strictly exactly one of the items in the " # noqa: E501f"categories given. Find a semantic match of category if possible. If it does not categorize well " # noqa: E501f"into any of the listed categories, categorize it as 'unknown'.\n\nText:\n\n{text}\n\n\nCategory:" # noqa: E501)settings_string = "".join(str(value) for value in settings.values())- cache_key = (- f"cache:{self.workflow_id}:{ToolUtils.hash_str(settings_string)}:"- f"{ToolUtils.hash_str(prompt)}"+ classification = self.helper.find_classification(+ use_cache=use_cache,+ settings_string=settings_string,+ prompt=prompt,+ bins=bins,+ llm=llm,)- classification = None- if use_cache:- cache = ToolCache(- tool=self,- platform_host=self.get_env_or_die(ToolEnv.PLATFORM_HOST),- platform_port=int(self.get_env_or_die(ToolEnv.PLATFORM_PORT)),- )- cached_response = cache.get(cache_key)- if cached_response is not None:- classification = cached_response- self.stream_cost(cost=0.0, cost_units="cache")-- if classification is None:- self.stream_log("Calling LLM")- try:- response = llm.complete(prompt, max_tokens=50, stop=["\n"])- if response is None:- self.stream_error_and_exit("Error calling LLM")- classification = response.text.strip()- self.stream_log(f"LLM response: {response}")- except Exception as e:- self.stream_error_and_exit(f"Error calling LLM: {e}")-- if use_cache:- cache = ToolCache(- tool=self,- platform_host=self.get_env_or_die(ToolEnv.PLATFORM_HOST),- platform_port=int(- self.get_env_or_die(ToolEnv.PLATFORM_PORT)- ),- )- cache.set(cache_key, classification)-- classification = classification.lower() # type: ignore- bins = [bin.lower() for bin in bins]-- if classification not in bins:- self.stream_error_and_exit(- f"Invalid classification done: {classification}"- )- if not classification:- classification = "unknown"-source_name = self.get_exec_metadata.get(MetadataKey.SOURCE_NAME)self._copy_input_to_output_bin(output_dir=output_dir,diff --git a/unstract/tool-registry/src/unstract/tool_registry/public_tools.json b/unstract/tool-registry/src/unstract/tool_registry/public_tools.jsonindex 904616f00..fe562f013 100644--- a/unstract/tool-registry/src/unstract/tool_registry/public_tools.json+++ b/unstract/tool-registry/src/unstract/tool_registry/public_tools.json@@ -5,7 +5,7 @@"schemaVersion": "0.0.1","displayName": "File Classifier","functionName": "classify",- "toolVersion": "0.0.1",+ "toolVersion": "0.0.2","description": "Classifies a file into a bin based on its contents","input": {"description": "File to be classified"@@ -36,6 +36,15 @@{"isEnabled": false}+ ],+ "textExtractors": [+ {+ "isEnabled": true,+ "adapterId": "textExtractorId",+ "title": "Text Extraction Adapter",+ "isRequired": false,+ "description": "Choose extractor adapter to extract text"+ }]},"ioCompatibility": {@@ -130,9 +139,9 @@}},"icon": "<?xml version=\"1.0\" encoding=\"UTF-8\" standalone=\"no\"?>\n<svg\n enable-background=\"new 0 0 20 20\"\n height=\"48\"\n viewBox=\"0 0 20 20\"\n width=\"48\"\n fill=\"#000000\"\n version=\"1.1\"\n id=\"svg8109\"\n sodipodi:docname=\"folder_copy_black_48dp.svg\"\n xmlns:inkscape=\"http://www.inkscape.org/namespaces/inkscape\"\n xmlns:sodipodi=\"http://sodipodi.sourceforge.net/DTD/sodipodi-0.dtd\"\n xmlns=\"http://www.w3.org/2000/svg\"\n xmlns:svg=\"http://www.w3.org/2000/svg\">\n <defs\n id=\"defs8113\" />\n <sodipodi:namedview\n id=\"namedview8111\"\n pagecolor=\"#ffffff\"\n bordercolor=\"#000000\"\n borderopacity=\"0.25\"\n inkscape:showpageshadow=\"2\"\n inkscape:pageopacity=\"0.0\"\n inkscape:pagecheckerboard=\"0\"\n inkscape:deskcolor=\"#d1d1d1\"\n showgrid=\"false\" />\n <g\n id=\"g8099\">\n <rect\n fill=\"none\"\n height=\"20\"\n width=\"20\"\n x=\"0\"\n id=\"rect8097\"\n y=\"0\" />\n </g>\n <g\n id=\"g8107\"\n style=\"fill:#ff4d6d;fill-opacity:1\">\n <g\n id=\"g8105\"\n style=\"fill:#ff4d6d;fill-opacity:1\">\n <path\n d=\"M 2.5,5 H 1 V 15.5 C 1,16.33 1.67,17 2.5,17 H 15.68 V 15.5 H 2.5 Z\"\n id=\"path8101\"\n style=\"fill:#ff4d6d;fill-opacity:1\" />\n <path\n d=\"M 16.5,4 H 11 L 9,2 H 5.5 C 4.67,2 4,2.67 4,3.5 v 9 C 4,13.33 4.67,14 5.5,14 h 11 c 0.83,0 1.5,-0.67 1.5,-1.5 v -7 C 18,4.67 17.33,4 16.5,4 Z m 0,8.5 h -11 v -9 h 2.88 l 2,2 h 6.12 z\"\n id=\"path8103\"\n style=\"fill:#ff4d6d;fill-opacity:1\" />\n </g>\n </g>\n</svg>\n",- "image_url": "docker:unstract/tool-classifier:0.0.1",+ "image_url": "docker:unstract/tool-classifier:0.0.2","image_name": "unstract/tool-classifier",- "image_tag": "0.0.1"+ "image_tag": "0.0.2"},"document_indexer": {"tool_uid": "document_indexer",