LLM-Driven Extraction of Unstructured Data — Built for API Deployments & ETL Pipeline Workflows
Python ai-agentsdata-engineeringdocument-aigenerative-aiidpjson-extractionllmmcp-serverocrpdf-extractionprompt-engineeringstructured-output
Fixed high token usage for json type (#397)
Deepak K42aa2152c97b
diff --git a/prompt-service/src/unstract/prompt_service/main.py b/prompt-service/src/unstract/prompt_service/main.pyindex decce974d..c2a1fc6ab 100644--- a/prompt-service/src/unstract/prompt_service/main.py+++ b/prompt-service/src/unstract/prompt_service/main.py@@ -497,24 +497,29 @@ def prompt_processor() -> Any:if answer.lower() == "[]" or answer.lower() == "na":structured_output[output[PSKeys.NAME]] = Noneelse:- prompt = f"Convert the following text into valid JSON string: \- \n{answer}\n\n The JSON string should be able to be parsed \- into a Python dictionary. \- Output just the JSON string. No explanation is required. \- If you cannot extract the JSON string, output {{}}"- answer = run_completion(- llm=llm,- prompt=prompt,- )try:structured_output[output[PSKeys.NAME]] = json.loads(answer)- except JSONDecodeError as e:- app.logger.info(f"JSON format error : {answer}", LogLevel.ERROR)- app.logger.info(- f"Error parsing response (to json): {e}",- LogLevel.ERROR,- )- structured_output[output[PSKeys.NAME]] = {}+ except JSONDecodeError:+ prompt = f"Convert the following text into valid JSON string: \+ \n{answer}\n\n The JSON string should be able to be parsed \+ into a Python dictionary. \+ Output just the JSON string. No explanation is required. \+ If you cannot extract the JSON string, output {{}}"+ try:+ answer = run_completion(+ llm=llm,+ prompt=prompt,+ )+ structured_output[output[PSKeys.NAME]] = json.loads(answer)+ except JSONDecodeError as e:+ app.logger.info(+ f"JSON format error : {answer}", LogLevel.ERROR+ )+ app.logger.info(+ f"Error parsing response (to json): {e}", LogLevel.ERROR+ )+ structured_output[output[PSKeys.NAME]] = {}+else:structured_output[output[PSKeys.NAME]] = answer@@ -546,6 +551,7 @@ def prompt_processor() -> Any:usage_kwargs=usage_kwargs,)challenge = challenge_plugin["entrypoint_cls"](+ llm=llm,challenge_llm=challenge_llm,run_id=run_id,context=context,