Skip to content

LLM-Driven Extraction of Unstructured Data — Built for API Deployments & ETL Pipeline Workflows

Fix/Subquestion token optimization (#655)

harini-venkataraman13f1dbf8ad75
diff --git a/prompt-service/src/unstract/prompt_service/main.py b/prompt-service/src/unstract/prompt_service/main.py
index 241c83b59..885a890b8 100644
--- a/prompt-service/src/unstract/prompt_service/main.py
+++ b/prompt-service/src/unstract/prompt_service/main.py
@@ -702,11 +702,20 @@ def run_retrieval( # type:ignore
"Generate set of specific subquestions "
"from the prompt which can be used to retrive "
"relevant context from vector db. "
+ "Use your logical abilities to "
+ " only generate as many subquestions as necessary "
+ " — fewer subquestions if the prompt is simpler. "
+ "Decide the minimum limit for subquestions "
+ "based on the complexity input prompt and set the maximum limit"
+ "for the subquestions to 10."
"Ensure that each subquestion is distinct and relevant"
- "to the different facets of the original query. Do not"
- "add subquestions for details not mentioned in the original "
- "prompt. The goal is to maximize retrieval accuracy"
- " using these subquestions. Please note that, there are cases where the "
+ "to the the original query. "
+ "Do not add subquestions for details"
+ "not mentioned in the original prompt."
+ " The goal is to maximize retrieval accuracy"
+ " using these subquestions. Use your logical abilities to ensure "
+ " that each subquestion targets a distinct aspect of the original query."
+ " Please note that, there are cases where the "
"response might have a list of answers. The subquestions must not miss out "
"any values in these cases. "
"Output should be a list of comma seperated "
@@ -718,6 +727,7 @@ def run_retrieval( # type:ignore
prompt=subq_prompt,
)
subquestion_list = subquestions.split(",")
+ raw_retrieved_context = ""
for each_subq in subquestion_list:
retrieved_context = _retrieve_context(
output, doc_id, vector_index, each_subq
@@ -726,8 +736,10 @@ def run_retrieval( # type:ignore
# inconsistency issue owing to risk of infinte loop
# and inablity to diffrentiate genuine cases of
# empty context.
-
- context = "".join([context, retrieved_context])
+ raw_retrieved_context = "\f\n".join(
+ [raw_retrieved_context, retrieved_context]
+ )
+ context = _remove_duplicate_nodes(raw_retrieved_context)
if retrieval_type == PSKeys.SIMPLE:
@@ -756,6 +768,12 @@ def run_retrieval( # type:ignore
return (answer, context)
+def _remove_duplicate_nodes(retrieved_context: str) -> str:
+ context_set: set[str] = set(retrieved_context.split("\f\n"))
+ fomatted_context = "\f\n".join(context_set)
+ return fomatted_context
+
+
def _retrieve_context(output, doc_id, vector_index, answer) -> str:
retriever = vector_index.as_retriever(
similarity_top_k=output[PSKeys.SIMILARITY_TOP_K],
@@ -774,7 +792,7 @@ def _retrieve_context(output, doc_id, vector_index, answer) -> str:
# ToDo: May have to fine-tune this value for node score or keep it
# configurable at the adapter level
if node.score > 0:
- text += node.get_content() + "\n"
+ text += node.get_content() + "\f\n"
else:
app.logger.info("Node score is less than 0. " f"Ignored: {node.score}")
return text