Granite RAG Library

Granite RAG Library

Adds six LoRA adapters for IBM Granite models that rewrite queries, ask for clarification, judge passages, check answerability and flag hallucinations in RAG pipelines. For developers building agentic retrieval.

Made by ibm-granite on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: The Granite RAG Library is a set of six LoRA adapters for IBM Granite language models. Each adapter does one job in a retrieval-augmented generation pipeline, such as rewriting a question or checking an answer against passages. They work with granite-4.0-micro and with the granite-4.1 models at 3B, 8B and 30B. Four of the adapters also come in aLoRA versions.

Who it suits: It suits developers who build RAG or agent pipelines and want small, task-specific helpers around a Granite model instead of one large prompt.

Industry | Who uses it | How they use it | Example
Software | RAG engineers | Add query rewriting before retrieval | Making a follow-up question stand alone
Support | Help desk assistant builders | Ask the user to clarify vague questions | Requesting detail when a query has two meanings
Enterprise search | Knowledge base teams | Judge whether retrieved passages are relevant | Dropping passages that would mislead the model
Quality assurance | AI evaluation teams | Flag unsupported sentences in generated answers | Scoring each sentence against its sources

Skip it if you do not use Granite models, since the adapters are built for them.

Where to use it:

• RAG and agent pipelines built around a Granite base model
• Before retrieval, between retrieval and generation, or after generation, depending on the adapter
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Query rewriting | Retrieval teams | A standalone version of the latest user question
Clarification | Chat designers | A request for detail when a question is unclear
Answerability check | Pipeline builders | A verdict on whether the passages can answer the question
Hallucination check | Evaluators | A risk range for each sentence of an answer

How to set it up: You need these first.

• Python with pip, to fetch the files
• One of the supported Granite base models, which each adapter sits on
• A runtime that can load LoRA adapters for that base model

1. Install the Hugging Face hub library with pip.
2. Download the repository files with the hf command.
3. Pick the adapter for your task and the matching base model.
4. Follow that adapter's own README for the exact steps.

pip install -U huggingface_hub
hf download ibm-granite/granitelib-rag-r1.0

How to use it: Each adapter takes a conversation between a user and an assistant, and most also take a set of passages. You choose the adapter that fits the stage of your pipeline. This card gives no run commands, since each adapter has its own README.

hf download ibm-granite/granitelib-rag-r1.0

Tool | What it does
Query Rewrite | Turns the last user question into a standalone one before retrieval
Query Clarification | Detects an unclear question and drafts a clarification request
Context Relevance | Labels a passage as relevant, partly relevant or irrelevant, or as misleading
Answerability Determination | Decides whether the passages can answer the question
Hallucination Detection | Gives a risk range per sentence of a generated answer

Example: A user asks a follow-up that only makes sense in context. Query Rewrite turns it into a full question, so the retriever finds better passages. After the model answers, Hallucination Detection scores each sentence of the answer against those passages.

For newcomers and experts: If you are new, think of these adapters as small checkers placed around a chatbot that answers from documents. Start with Query Rewrite, since it is the first step and easiest to see working.

If you know the field, the adapters sit at three points. Query Rewrite is pre-retrieval. Query Clarification works before retrieval or before generation. Context Relevance and Answerability are pre-generation, and Hallucination Detection is post-generation. Context Relevance is released only for granite-4.0-micro.

User question
Query Rewrite makes it standalone
Retriever fetches passages
Relevance and answerability checks
Model answers, then hallucination check

Good to know:

• Base models: granite-4.0-micro, plus granite-4.1 at 3B, 8B and 30B
• Variants: aLoRA versions exist for four of the adapters
• Formats: Safetensors and GGUF
• Detail: each adapter has its own README with the specifics

• License: Apache 2.0
• Cost to run: none named, the source lists no API key, account or paid service
• Data it reaches: the conversations and passages you pass to your own pipeline, and the source names no web or account access
• Not for: use without a Granite base model, since the adapters are made for those

On Hugging Face, 13 November 2026

• Library: granite-library
• Parameters: 14.4M
• Weights: about 7 MB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 10,360
• Likes: 62
• License: apache-2.0
• Languages: en
• Based on: ibm-granite/granite-4.0-micro
• Tags: granite-library, safetensors, gguf

Links

• Model: https://huggingface.co/ibm-granite/granitelib-rag-r1.0
• Files: https://huggingface.co/ibm-granite/granitelib-rag-r1.0/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ibm-granite/granitelib-rag-r1.0

Then follow the usage in its model card. Nothing in this kit runs by itself.