Skip to content

Memory RAG fails to upload PDF – "model granite-embedding-107m-multilingual does not exist" when using custom embedding model #9307

@zsj1029

Description

@zsj1029

LocalAI version:

LocalAI v4.1.3 (fdc9f7b)

Environment, CPU architecture, OS, and Version:

Linux localai 5.15.0-173-generic #183~20.04.1-Ubuntu SMP Fri Mar 13 14:51:03 UTC 2026 x86_64 x86_64 x86_64 GNU/Linux

Describe the bug
Description

I configured an Embedding Model in the Web system settings.

Image

Following the official documentation, I downloaded and configured the following model:
Qwen3-Embedding-8B-Q4_K_M.ggufMy configuration is as follows:

YAML

backend: llama-cpp
embeddings: true
min_p: 0
name: granite-embedding-107m-multilingual
options:
  - use_jinja: true
parameters:
  min_p: 0
  model: Qwen3-Embedding-8B-Q4_K_M.gguf
  repeat_penalty: 1
  temperature: 0.6
  top_k: 20
  top_p: 0.95
repeat_penalty: 1
temperature: 0.6
template:
  use_tokenizer_template: true
top_k: 20
top_p: 0.95

When using Memory RAG to upload a PDF file, the POST /embedding request fails with the error:
model granite-embedding-107m-multilingual does not exist.

After checking the source code, it seems that LocalAI requires a default embedding model to be configured at container startup. Otherwise, it always falls back to granite-embedding-107m-multilingual.

Problem
I couldn't find any clear way to set a custom default embedding model (unlike Open WebUI, where you can easily specify both LLM and embedding model for RAG).
I want to use a custom LLM + custom embedding model for RAG, but it feels unnecessarily difficult in LocalAI.

Expected Behavior
LocalAI should allow flexible specification of the embedding model name — whether through the name field in the YAML config, environment variables, Web UI settings, or startup parameters — without being forced to use or fall back to granite-embedding-107m-multilingual.

Actual Behavior
Even after configuring a custom embedding model (including changing the name field), RAG functionality still tries to call the non-existent granite-embedding-107m-multilingual model, causing failure.

Logs

Additional context

Thanks , Great Product

Metadata

Metadata

Assignees

No one assigned

    Labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions