From 9f2fb4611fc8734238c188294b5f21d821583bf9 Mon Sep 17 00:00:00 2001
From: Rene Arredondo <120709323+Rene0422@users.noreply.github.com>
Date: Sun, 17 May 2026 23:11:54 -0700
Subject: [PATCH] Fix: guard empty/whitespace embedding inputs in LLMBundle
 (#14428) (#14924)

Closes #14428


### Type of change

- [x] Bug Fix (non-breaking change which fixes an issue)
---
 api/db/services/llm_service.py | 27 ++++++++++++++++++++++++++-
 1 file changed, 26 insertions(+), 1 deletion(-)

diff --git a/api/db/services/llm_service.py b/api/db/services/llm_service.py
index 60090bb040..9b6b5bd4f1 100644
--- a/api/db/services/llm_service.py
+++ b/api/db/services/llm_service.py
@@ -97,7 +97,24 @@ class LLMBundle(LLM4Tenant):
             generation = self.langfuse.start_observation(trace_context=self.trace_context, as_type="generation", name="encode", model=self.model_config["llm_name"], input={"texts": texts})
 
         safe_texts = []
-        for text in texts:
+        for idx, text in enumerate(texts):
+            # Embedding APIs (OpenAI-compatible, Zhipu, etc.) reject empty or
+            # whitespace-only inputs with errors like "Input at index N cannot
+            # be empty or whitespace only". Upstream parsers can produce such
+            # chunks — e.g. when OCR/vision on an embedded DOCX image returns
+            # nothing, or a table has only empty cells — so coerce to a safe
+            # placeholder here, at the single boundary every embedding path
+            # funnels through.
+            if text is None or not str(text).strip():
+                marker = "None" if text is None else "whitespace-only"
+                logging.warning(
+                    "LLMBundle.encode: empty input at index %d (%s) coerced to placeholder 'None' for model %s",
+                    idx,
+                    marker,
+                    self.model_config["llm_name"],
+                )
+                safe_texts.append("None")
+                continue
             token_size = num_tokens_from_string(text)
             if token_size > self.max_length:
                 target_len = int(self.max_length * 0.95)
@@ -121,6 +138,14 @@ class LLMBundle(LLM4Tenant):
         if self.langfuse:
             generation = self.langfuse.start_observation(trace_context=self.trace_context, as_type="generation", name="encode_queries", model=self.model_config["llm_name"], input={"query": query})
 
+        if query is None or not str(query).strip():
+            marker = "None" if query is None else "whitespace-only"
+            logging.warning(
+                "LLMBundle.encode_queries: empty query (%s) coerced to placeholder 'None' for model %s",
+                marker,
+                self.model_config["llm_name"],
+            )
+            query = "None"
         emd, used_tokens = self.mdl.encode_queries(query)
         if self.model_config["llm_factory"] == "Builtin":
             logging.info("LLMBundle.encode_queries query: {}, emd len: {}, used_tokens: {}. Builtin model don't need to update token usage".format(query, len(emd), used_tokens))