LLM-Driven Extraction of Unstructured Data — Built for API Deployments & ETL Pipeline Workflows
Python ai-agentsdata-engineeringdocument-aigenerative-aiidpjson-extractionllmmcp-serverocrpdf-extractionprompt-engineeringstructured-output
[FIX] Prompt service v2 fixes (#1247)
harini-venkataraman0bfc37f5b984
diff --git a/backend/prompt_studio/prompt_studio_core_v2/prompt_studio_helper.py b/backend/prompt_studio/prompt_studio_core_v2/prompt_studio_helper.pyindex fe045df49..f36a08500 100644--- a/backend/prompt_studio/prompt_studio_core_v2/prompt_studio_helper.py+++ b/backend/prompt_studio/prompt_studio_core_v2/prompt_studio_helper.py@@ -859,25 +859,6 @@ class PromptStudioHelper:doc_path = str(doc_path.parent.parent / "summarize" / (doc_path.stem + ".txt"))- summarize_file_path = PromptStudioHelper.summarize(- filename, org_id, document_id, is_summary, run_id, tool, doc_id- )- summarize_doc_id = IndexingUtils.generate_index_key(- vector_db=str(default_profile.vector_store.id),- embedding=str(default_profile.embedding_model.id),- x2text=str(default_profile.x2text.id),- chunk_size=str(default_profile.chunk_size),- chunk_overlap=str(default_profile.chunk_overlap),- file_path=summarize_file_path,- fs=fs_instance,- tool=util,- )- PromptStudioIndexHelper.handle_index_manager(- document_id=document_id,- is_summary=is_summary,- profile_manager=profile_manager,- doc_id=summarize_doc_id,- )logger.info("Summary enabled, set chunk to zero..")logger.info(f"Indexing document {doc_path} for {doc_id}")index_result = PromptStudioHelper.dynamic_indexer(diff --git a/backend/sample.env b/backend/sample.envindex e99074eab..102fb92d8 100644--- a/backend/sample.env+++ b/backend/sample.env@@ -75,9 +75,9 @@ PROMPT_STUDIO_FILE_PATH=/app/prompt-studio-data# Structure Tool Image (Runs prompt studio exported tools)# https://hub.docker.com/r/unstract/tool-structure-STRUCTURE_TOOL_IMAGE_URL="docker:unstract/tool-structure:0.0.73"+STRUCTURE_TOOL_IMAGE_URL="docker:unstract/tool-structure:0.0.74"STRUCTURE_TOOL_IMAGE_NAME="unstract/tool-structure"-STRUCTURE_TOOL_IMAGE_TAG="0.0.73"+STRUCTURE_TOOL_IMAGE_TAG="0.0.74"# Feature FlagsEVALUATION_SERVER_IP=unstract-fliptdiff --git a/prompt-service/src/unstract/prompt_service/controllers/answer_prompt.py b/prompt-service/src/unstract/prompt_service/controllers/answer_prompt.pyindex 1184f9e45..f620e24e6 100644--- a/prompt-service/src/unstract/prompt_service/controllers/answer_prompt.py+++ b/prompt-service/src/unstract/prompt_service/controllers/answer_prompt.py@@ -64,6 +64,7 @@ def prompt_processor() -> Any:variable_names: list[str] = []# Identifier for source of invocationexecution_source = payload.get(PSKeys.EXECUTION_SOURCE, "")+ context_retrieval_metrics = {}publish_log(log_events_id,{"tool_id": tool_id, "run_id": run_id, "doc_name": doc_name},@@ -272,6 +273,7 @@ def prompt_processor() -> Any:chunk_size=chunk_size,execution_source=execution_source,file_path=file_path,+ context_retrieval_metrics=context_retrieval_metrics,)metadata[PSKeys.CONTEXT][output[PSKeys.NAME]] = contextelse:@@ -568,7 +570,7 @@ def prompt_processor() -> Any:)metrics.setdefault(prompt_name, {}).update({- "context_retrieval": index.get_metrics(),+ "context_retrieval": context_retrieval_metrics.get(prompt_name, {}),f"{llm.get_usage_reason()}_llm": llm.get_metrics(),**challenge_metrics,}diff --git a/prompt-service/src/unstract/prompt_service/services/answer_prompt.py b/prompt-service/src/unstract/prompt_service/services/answer_prompt.pyindex a4b703e3e..aee5cf1f6 100644--- a/prompt-service/src/unstract/prompt_service/services/answer_prompt.py+++ b/prompt-service/src/unstract/prompt_service/services/answer_prompt.py@@ -1,4 +1,3 @@-import osfrom logging import Loggerfrom typing import Any, Optional@@ -228,12 +227,6 @@ class AnswerPromptService:raise APIError(PSKeys.PAID_FEATURE_MSG)extract_file_path = file_path- if execution_source == ExecutionSource.IDE.value:- # Adjust file path to read from the extract folder- base_name = os.path.splitext(os.path.basename(file_path))[0]- extract_file_path = os.path.join(- os.path.dirname(file_path), "extract", f"{base_name}.txt"- )# Read file content into contextfs_instance: FileStorage = FileStorage(FileStorageProvider.LOCAL)diff --git a/prompt-service/src/unstract/prompt_service/services/extraction.py b/prompt-service/src/unstract/prompt_service/services/extraction.pyindex 0793d5ca6..6d188425f 100644--- a/prompt-service/src/unstract/prompt_service/services/extraction.py+++ b/prompt-service/src/unstract/prompt_service/services/extraction.py@@ -1,3 +1,4 @@+from pathlib import Pathfrom typing import Any, Optionalfrom unstract.prompt_service.constants import ExecutionSource@@ -85,7 +86,7 @@ class ExtractionService:metadata = {X2TextConstants.WHISPER_HASH: whisper_hash_value}for key, value in metadata.items():tool_exec_metadata[key] = value- metadata_path = execution_run_data_folder / IKeys.METADATA_FILE+ metadata_path = str(Path(execution_run_data_folder / IKeys.METADATA_FILE))ToolUtils.dump_json(file_to_dump=metadata_path,json_to_dump=metadata,diff --git a/prompt-service/src/unstract/prompt_service/services/retrieval.py b/prompt-service/src/unstract/prompt_service/services/retrieval.pyindex 1e6cff78b..3dc6ea364 100644--- a/prompt-service/src/unstract/prompt_service/services/retrieval.py+++ b/prompt-service/src/unstract/prompt_service/services/retrieval.py@@ -1,3 +1,4 @@+import datetimefrom typing import Anyfrom unstract.prompt_service.constants import PromptServiceConstants as PSKeys@@ -5,6 +6,7 @@ from unstract.prompt_service.core.retrievers.simple import SimpleRetrieverfrom unstract.prompt_service.core.retrievers.subquestion import SubquestionRetrieverfrom unstract.prompt_service.services.answer_prompt import AnswerPromptServicefrom unstract.prompt_service.utils.file_utils import FileUtils+from unstract.prompt_service.utils.metrics import Metricsfrom unstract.sdk.llm import LLMfrom unstract.sdk.vector_db import VectorDB@@ -22,11 +24,15 @@ class RetrievalService:chunk_size: int,execution_source: str,file_path: str,+ context_retrieval_metrics: dict[str, Any],) -> tuple[str, list[str]]:context: list[str]if chunk_size == 0:context = RetrievalService.retrieve_complete_context(- execution_source=execution_source, file_path=file_path+ execution_source=execution_source,+ file_path=file_path,+ context_retrieval_metrics=context_retrieval_metrics,+ prompt_key=output[PSKeys.PROMPTX],)else:context = RetrievalService.run_retrieval(@@ -35,6 +41,7 @@ class RetrievalService:llm=llm,vector_db=vector_db,retrieval_type=retrieval_type,+ context_retrieval_metrics=context_retrieval_metrics,)answer = AnswerPromptService.construct_and_run_prompt( # type:ignoretool_settings=tool_settings,@@ -55,10 +62,13 @@ class RetrievalService:llm: LLM,vector_db: VectorDB,retrieval_type: str,+ context_retrieval_metrics: dict[str, Any],) -> list[str]:context: set[str]prompt = output[PSKeys.PROMPTX]top_k = output[PSKeys.SIMILARITY_TOP_K]+ prompt_key = output[PSKeys.NAME]+ retrieval_start_time = datetime.datetime.now()if retrieval_type == PSKeys.SUBQUESTION:context = SubquestionRetriever(vector_db=vector_db,@@ -75,11 +85,18 @@ class RetrievalService:top_k=top_k,llm=llm,).retrieve()-+ context_retrieval_metrics[prompt_key] = {+ "time_taken(s)": Metrics.elapsed_time(start_time=retrieval_start_time)+ }return list(context)@staticmethod- def retrieve_complete_context(execution_source: str, file_path: str) -> list[str]:+ def retrieve_complete_context(+ execution_source: str,+ file_path: str,+ context_retrieval_metrics: dict[str, Any],+ prompt_key: str,+ ) -> list[str]:"""Loads full context from raw file for zero chunk size retrievalArgs:@@ -90,5 +107,9 @@ class RetrievalService:list[str]: context from extracted file."""fs_instance = FileUtils.get_fs_instance(execution_source=execution_source)+ retrieval_start_time = datetime.datetime.now()context = fs_instance.read(path=file_path, mode="r")+ context_retrieval_metrics[prompt_key] = {+ "time_taken(s)": Metrics.elapsed_time(start_time=retrieval_start_time)+ }return [context]diff --git a/prompt-service/src/unstract/prompt_service/utils/metrics.py b/prompt-service/src/unstract/prompt_service/utils/metrics.pynew file mode 100644index 000000000..02a5783f4--- /dev/null+++ b/prompt-service/src/unstract/prompt_service/utils/metrics.py@@ -0,0 +1,9 @@+import datetime+++class Metrics:++ @staticmethod+ def elapsed_time(start_time) -> float:+ """Returns the elapsed time since the process was started."""+ return (datetime.datetime.now() - start_time).total_seconds()diff --git a/tools/structure/src/helpers.py b/tools/structure/src/helpers.pyindex e29b5f31c..aac6a4293 100644--- a/tools/structure/src/helpers.py+++ b/tools/structure/src/helpers.py@@ -1,3 +1,4 @@+import datetimefrom typing import Any, Optionalfrom constants import IndexingConstants as IKeys@@ -91,3 +92,8 @@ class StructureToolHelper:)doc_id = responder.index(payload=payload)return doc_id++ @staticmethod+ def elapsed_time(start_time) -> float:+ """Returns the elapsed time since the process was started."""+ return (datetime.datetime.now() - start_time).total_seconds()diff --git a/tools/structure/src/main.py b/tools/structure/src/main.pyindex 34e52b9f1..0f0d79492 100644--- a/tools/structure/src/main.py+++ b/tools/structure/src/main.py@@ -1,3 +1,4 @@+import datetimeimport jsonimport loggingimport os@@ -131,7 +132,6 @@ class StructureTool(BaseTool):if tool_settings[SettingsKeys.ENABLE_SINGLE_PASS_EXTRACTION]:if summarize_as_source:summarize_file_hash = self._summarize_and_index(- tool_id=tool_id,tool_settings=tool_settings,tool_data_dir=tool_data_dir,responder=responder,@@ -139,25 +139,9 @@ class StructureTool(BaseTool):usage_kwargs=usage_kwargs,)payload[SettingsKeys.FILE_HASH] = summarize_file_hash- # STHelper.dynamic_indexing(- # tool_settings=tool_settings,- # run_id=self.file_execution_id,- # file_path=tool_data_dir / SettingsKeys.EXTRACT,- # tool=self,- # execution_run_data_folder=execution_run_data_folder,- # chunk_overlap=tool_settings[SettingsKeys.CHUNK_OVERLAP],- # chunk_size=tool_settings[SettingsKeys.CHUNK_SIZE],- # reIndex=True,- # usage_kwargs=usage_kwargs,- # enable_highlight=enable_highlight,- # extracted_text=extracted_text,- # tool_id=tool_metadata[SettingsKeys.TOOL_ID],- # file_hash=file_hash,- # )- # Handle metrics for single pass extraction- # index_metrics = {SettingsKeys.INDEXING: index.get_metrics()}-self.stream_log("Fetching response for single pass extraction")+ # Since indexing is not involved for single pass+ index_metrics = {"time_taken(s)": 0}prompt_service_resp = responder.single_pass_extraction(payload=payload,)@@ -169,7 +153,6 @@ class StructureTool(BaseTool):for output in outputs:if summarize_as_source:summarize_file_hash = self._summarize_and_index(- tool_id=tool_id,tool_settings=tool_settings,tool_data_dir=tool_data_dir,responder=responder,@@ -178,6 +161,8 @@ class StructureTool(BaseTool):)payload[SettingsKeys.OUTPUTS] = outputspayload[SettingsKeys.FILE_HASH] = summarize_file_hash+ # Since indexing is not involved for summary+ index_metrics[output[SettingsKeys.NAME]] = {"time_taken(s)": 0}breakself.stream_log(f"Chunk size '{output[SettingsKeys.CHUNK_SIZE]}', "@@ -186,6 +171,7 @@ class StructureTool(BaseTool):if (reindex or not summarize_as_source) and output[SettingsKeys.CHUNK_SIZE] != 0:+ indexing_start_time = datetime.datetime.now()self.stream_log(f"Sucessfully extracted text, "f"indexing {tool_data_dir / SettingsKeys.EXTRACT}.."@@ -205,11 +191,13 @@ class StructureTool(BaseTool):file_hash=file_hash,extracted_text=extracted_text,)- # Handle metrics for single pass extraction- # index_metrics[output[SettingsKeys.NAME]] = {- # SettingsKeys.INDEXING: index.get_metrics()- # }- # index.clear_metrics()+ index_metrics[output[SettingsKeys.NAME]] = {+ SettingsKeys.INDEXING: {+ "time_taken(s)": STHelper.elapsed_time(+ start_time=indexing_start_time+ )+ }+ }reindex = Falseexcept Exception as e:@@ -289,14 +277,11 @@ class StructureTool(BaseTool):def _summarize_and_index(self,- tool_id: str,tool_settings: dict[str, Any],tool_data_dir: Path,responder: PromptTool,outputs: dict[str, Any],usage_kwargs: dict[Any, Any] = {},- enable_highlight: bool = False,- execution_run_data_folder: str = None,) -> str:"""Summarizes the context of the file and indexes the summarizedcontent.@@ -364,21 +349,6 @@ class StructureTool(BaseTool):summarize_file_hash: str = self.workflow_filestorage.get_hash_from_file(path=summarize_file_path)- STHelper.dynamic_indexing(- file_path=summarize_file_path,- tool_settings=tool_settings,- run_id=self.file_execution_id,- execution_run_data_folder=execution_run_data_folder,- reindex=True,- usage_kwargs=usage_kwargs,- enable_highlight=enable_highlight,- file_hash=summarize_file_hash,- tool_id=tool_id,- tool=self,- chunk_overlap=0,- chunk_size=0,- extracted_text=summarized_context,- )return summarize_file_hash