图片知识库报错code=105000030 message=parser parse failed: parse document failed, err: [ParseImage] model generate failed: error during Chat request: 400 Bad Request: illegal base64 data at input byte 4

文本知识库无问题，图片知识库报下面错误：
code=105000030 message=parser parse failed: parse document failed, err: [ParseImage] model generate failed: error during Chat request: 400 Bad Request: illegal base64 data at input byte 4

关键配置：
# Settings for Model
# Model for agent & workflow
# add suffix number to add different models
export MODEL_PROTOCOL_0="ollama" # protocol
export MODEL_OPENCOZE_ID_0="100001" # id for record
export MODEL_NAME_0="llava:7b" # model name for show
export MODEL_ID_0="llava:7b" # model name for connection
export MODEL_API_KEY_0="" # model api key
export MODEL_BASE_URL_0="http://host.docker.internal:11434" # model base url

# Model for knowledge nl2sql, messages2query (rewrite), image annotation, workflow knowledge recall
# add prefix to assign specific model, downgrade to default config when prefix is not configured:
# 1. nl2sql: NL2SQL_ (e.g. NL2SQL_BUILTIN_CM_TYPE)
# 2. messages2query: M2Q_ (e.g. M2Q_BUILTIN_CM_TYPE)
# 3. image annotation: IA_ (e.g. IA_BUILTIN_CM_TYPE)
# 4. workflow knowledge recall: WKR_ (e.g. WKR_BUILTIN_CM_TYPE)
# supported chat model type: openai / ark / deepseek / ollama / qwen / gemini
export BUILTIN_CM_TYPE="ollama"

# 显式指定图像标注任务使用 Ollama
export IA_BUILTIN_CM_TYPE="ollama"
export IA_BUILTIN_CM_OLLAMA_BASE_URL="http://host.docker.internal:11434" # 替换为您自己的宿主机IP
export IA_BUILTIN_CM_OLLAMA_MODEL="llava:7b"

# type ollama
export BUILTIN_CM_OLLAMA_BASE_URL="http://host.docker.internal:11434"
export BUILTIN_CM_OLLAMA_MODEL="llava:7b"

# Settings for Embedding
# The Embedding model relied on by knowledge base vectorization does not need to be configured
# if the vector database comes with built-in Embedding functionality (such as VikingDB). Currently,
# Coze Studio supports four access methods: openai, ark, ollama, and custom http. Users can simply choose one of them when using
# embedding type: ark / openai / ollama / gemini / http
export EMBEDDING_TYPE="ollama"
export EMBEDDING_MAX_BATCH_SIZE=100

# ollama embedding
export OLLAMA_EMBEDDING_BASE_URL="http://host.docker.internal:11434" # (string, required) Ollama embedding base_url
export OLLAMA_EMBEDDING_MODEL="bge-m3" # (string, required) Ollama embedding model
export OLLAMA_EMBEDDING_DIMS="1024" # (int, required) Ollama embedding dimensions
/

The payload names the ParseImage path and the IA_BUILTIN_CM_OLLAMA configuration, but no source file or test. Start by reproducing image knowledge-base ingestion with the shown Ollama settings and trace the failing Chat request; done means the image document parses successfully without the illegal-base64 error.

Written by the indexing model from the issue text.