Skip to content

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

UN-2967 [FIX] - Fix Repeated Embedding Call For Zero Chunk Size and Multiple Prompts (#1682)

Praveen Kumar0e834f57f8ae
diff --git a/prompt-service/src/unstract/prompt_service/controllers/answer_prompt.py b/prompt-service/src/unstract/prompt_service/controllers/answer_prompt.py
index 010637a86..1ff223eeb 100644
--- a/prompt-service/src/unstract/prompt_service/controllers/answer_prompt.py
+++ b/prompt-service/src/unstract/prompt_service/controllers/answer_prompt.py
@@ -149,19 +149,24 @@ def prompt_processor() -> Any:
capture_metrics=True,
)
- embedding = EmbeddingCompat(
- adapter_instance_id=output[PSKeys.EMBEDDING],
- tool=util,
- kwargs={
- **usage_kwargs,
- },
- )
+ # Only create embedding and vector_db if chunk_size > 0
+ # When chunk_size is 0, we read the complete file without embeddings
+ embedding = None
+ vector_db = None
+ if chunk_size > 0:
+ embedding = EmbeddingCompat(
+ adapter_instance_id=output[PSKeys.EMBEDDING],
+ tool=util,
+ kwargs={
+ **usage_kwargs,
+ },
+ )
- vector_db = VectorDB(
- tool=util,
- adapter_instance_id=output[PSKeys.VECTOR_DB],
- embedding=embedding,
- )
+ vector_db = VectorDB(
+ tool=util,
+ adapter_instance_id=output[PSKeys.VECTOR_DB],
+ embedding=embedding,
+ )
except SdkError as e:
msg = f"Couldn't fetch adapter. {e}"
app.logger.error(msg)
@@ -384,7 +389,7 @@ def prompt_processor() -> Any:
output=output,
doc_id=doc_id,
llm=llm,
- vector_db=vector_db,
+ vector_db=vector_db, # This will be None when chunk_size is 0
retrieval_type=retrieval_strategy,
metadata=metadata,
chunk_size=chunk_size,
@@ -657,7 +662,9 @@ def prompt_processor() -> Any:
**challenge_metrics,
}
)
- vector_db.close()
+ # Only close vector_db if it was created (chunk_size > 0)
+ if vector_db:
+ vector_db.close()
publish_log(
log_events_id,
{"tool_id": tool_id, "doc_name": doc_name},
diff --git a/prompt-service/src/unstract/prompt_service/services/indexing.py b/prompt-service/src/unstract/prompt_service/services/indexing.py
index 02cf65512..b71a4f989 100644
--- a/prompt-service/src/unstract/prompt_service/services/indexing.py
+++ b/prompt-service/src/unstract/prompt_service/services/indexing.py
@@ -51,6 +51,13 @@ class IndexingService:
tool=util,
fs=fs_instance,
)
+
+ # Skip embedding creation and indexing when chunk_size is 0
+ # When chunk_size is 0, we don't need vector operations
+ if chunking_config.chunk_size == 0:
+ logger.info(f"Skipping indexing for chunk_size=0. Doc ID: {doc_id}")
+ return doc_id
+
embedding = EmbeddingCompat(
adapter_instance_id=instance_identifiers.embedding_instance_id,
tool=util,