This is an interactive notebook. You can run it locally or use the links below:

​
Prerequisites

First, install the necessary libraries, set up API keys, log in to W&B, and create a new W&B project.
  1. Install weave, pandas, unsloth, wandb, litellm, pydantic, torch, and faiss-gpu using pip.
%%capture
!pip install weave wandb pandas pydantic litellm faiss-gpu
python
%%capture
!pip install unsloth
# Also get the latest nightly Unsloth!
!pip uninstall unsloth -y && pip install --upgrade --no-cache-dir "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
  1. Add the necessary API keys from your environment.
import os

from google.colab import userdata

os.environ["WANDB_API_KEY"] = userdata.get("WANDB_API_KEY")  # W&B Models and Weave
os.environ["OPENAI_API_KEY"] = userdata.get(
    "OPENAI_API_KEY"
)  # OpenAI - for retrieval embeddings
os.environ["GEMINI_API_KEY"] = userdata.get(
    "GEMINI_API_KEY"
)  # Gemini - for the base chat model
  1. Log in to W&B, and create a new project.
import pandas as pd
import wandb

import weave

wandb.login()

PROJECT = "weave-cookboook-demo"
ENTITY = "wandb-smle"

weave.init(ENTITY + "/" + PROJECT)

​
Download ChatModel from Models Registry and implement UnslothLoRAChatModel

In our scenario, the Llama-3.2 model has already been fine-tuned by the Model Team using the unsloth library for performance optimization, and is available in the W&B Models Registry. In this step, we’ll retrieve the fine-tuned ChatModel from the Registry and convert it into a weave.Model to make it compatible with the RagModel.
The RagModel referenced below is a top-level weave.Model that can be considered a complete RAG Application. It contains a ChatModel, vector database, and a prompt. The ChatModel is also a weave.Model, which contains code to download an artifact from the W&B Registry. ChatModel can be changed modularly to support any kind of other LLM chat model as part of the RagModel. For more information, view the model in Weave.
To load the ChatModel, unsloth.FastLanguageModel or peft.AutoPeftModelForCausalLM with adapters are used, enabling efficient integration into the app. After downloading the model from the Registry, you can set up the initialization and prediction logic by using the model_post_init method. The required code for this step is available in the Use tab of the Registry and can be copied directly into your implementation The code below defines the UnslothLoRAChatModel class to manage, initialize, and use the fine-tuned Llama-3.2 model retrieved from the W&B Models Registry. UnslothLoRAChatModel uses unsloth.FastLanguageModel for optimized inference. The model_post_init method handles downloading and setting up the model, while the predict method processes user queries and generates responses. To adapt the code for your use case, update the MODEL_REG_URL with the correct Registry path for your fine-tuned model and adjust parameters like max_seq_length or dtype based on your hardware or requirements. 
from typing import Any

from pydantic import PrivateAttr
from unsloth import FastLanguageModel

import weave

class UnslothLoRAChatModel(weave.Model):
    """
    We define an extra ChatModel class to be able store and version more parameters than just the model name.
    Especially, relevant if we consider fine-tuning (locally or aaS) because of specific parameters.
    """

    chat_model: str
    cm_temperature: float
    cm_max_new_tokens: int
    cm_quantize: bool
    inference_batch_size: int
    dtype: Any
    device: str
    _model: Any = PrivateAttr()
    _tokenizer: Any = PrivateAttr()

    def model_post_init(self, __context):
        # we can simply paste this from the "Use" tab from the registry
        run = wandb.init(project=PROJECT, job_type="model_download")
        artifact = run.use_artifact(f"{self.chat_model}")
        model_path = artifact.download()

        # unsloth version (enable native 2x faster inference)
        self._model, self._tokenizer = FastLanguageModel.from_pretrained(
            model_name=model_path,
            max_seq_length=self.cm_max_new_tokens,
            dtype=self.dtype,
            load_in_4bit=self.cm_quantize,
        )
        FastLanguageModel.for_inference(self._model)

    @weave.op()
    async def predict(self, query: list[str]) -> dict:
        # add_generation_prompt = true - Must add for generation
        input_ids = self._tokenizer.apply_chat_template(
            query,
            tokenize=True,
            add_generation_prompt=True,
            return_tensors="pt",
        ).to("cuda")

        output_ids = self._model.generate(
            input_ids=input_ids,
            max_new_tokens=64,
            use_cache=True,
            temperature=1.5,
            min_p=0.1,
        )

        decoded_outputs = self._tokenizer.batch_decode(
            output_ids[0][input_ids.shape[1] :], skip_special_tokens=True
        )

        return "".join(decoded_outputs).strip()
python
MODEL_REG_URL = "wandb32/wandb-registry-RAG Chat Models/Finetuned Llama-3.2:v3"

max_seq_length = 2048  # Choose any! We auto support RoPE Scaling internally!
dtype = (
    None  # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
)
load_in_4bit = True  # Use 4bit quantization to reduce memory usage. Can be False.

new_chat_model = UnslothLoRAChatModel(
    name="UnslothLoRAChatModelRag",
    chat_model=MODEL_REG_URL,
    cm_temperature=1.0,
    cm_max_new_tokens=max_seq_length,
    cm_quantize=load_in_4bit,
    inference_batch_size=max_seq_length,
    dtype=dtype,
    device="auto",
)
python
await new_chat_model.predict(
    [{"role": "user", "content": "What is the capital of Germany?"}]
)

​
Integrate the new ChatModel version into RagModel

Building a RAG application from a fine-tuned chat model improves conversational AI by using tailored components without having to rebuild the entire pipeline. In this step, we retrieve the existing RagModel from our Weave project and update its ChatModel to use the newly fine-tuned model. This seamless swap means that other components like the vector database (VDB) and prompts remain untouched, preserving the application’s overall structure while improving performance. The code below retrieves the RagModel object using a reference from the Weave project. The chat_model attribute of the RagModel is then updated to use the new UnslothLoRAChatModel instance created in the previous step. After this, the updated RagModel is published to create a new version. Finally, the updated RagModel is used to run a sample prediction query, verifying that the new chat model is being used. 
RagModel = weave.ref(
    "weave:///wandb-smle/weave-cookboook-demo/object/RagModel:cqRaGKcxutBWXyM0fCGTR1Yk2mISLsNari4wlGTwERo"
).get()
python
RagModel.chat_model.chat_model
python
await RagModel.predict("When was the first conference on climate change?")
python
# MAGIC: exchange chat_model and publish new version (no need to worry about other RAG components)
RagModel.chat_model = new_chat_model
python
RagModel.chat_model.chat_model
python
# first publish new version so that in prediction we reference new version
PUB_REFERENCE = weave.publish(RagModel, "RagModel")
python
await RagModel.predict("When was the first conference on climate change?")

​
Run a weave.Evaluation

In the next step, we evaluate the performance of our updated RagModel using an existing weave.Evaluation. This process ensures that the new fine-tuned chat model is performing as expected within the RAG application. To streamline integration and enable collaboration between the Models and Apps teams, we log evaluation results for both the model’s W&B run and as part of the Weave workspace. In Models:
  • The evaluation summary is logged to the W&B run used to download the fine-tuned chat model. This includes summary metrics and graphs displayed in a workspace view for analysis.
  • The evaluation trace ID is added to the run’s configuration, linking directly to the Weave page for easier traceability by the Model Team.
In Weave:
  • The artifact or registry link for the ChatModel is stored as an input to the RagModel.
  • The W&B run ID is saved as an extra column in the evaluation traces for better context.
The code below demonstrates how to retrieve an evaluation object, execute the evaluation using the updated RagModel, and log the results to both W&B and Weave. Ensure that the evaluation reference (WEAVE_EVAL) matches your project setup. 
# MAGIC: we can simply get an evaluation with a eval dataset and scorers and use them
WEAVE_EVAL = "weave:///wandb-smle/weave-cookboook-demo/object/climate_rag_eval:ntRX6qn3Tx6w3UEVZXdhIh1BWGh7uXcQpOQnIuvnSgo"
climate_rag_eval = weave.ref(WEAVE_EVAL).get()
python
with weave.attributes({"wandb-run-id": wandb.run.id}):
    # use .call attribute to retrieve both the result and the call in order to save eval trace to Models
    summary, call = await climate_rag_eval.evaluate.call(climate_rag_eval, RagModel)
python
# log to models
wandb.run.log(pd.json_normalize(summary, sep="/").to_dict(orient="records")[0])
wandb.run.config.update(
    {"weave_url": f"https://wandb.ai/wandb-smle/weave-cookboook-demo/r/call/{call.id}"}
)
wandb.run.finish()

​
Save the new RAG Model to the Registry

To make the updated RagModel available for future use by both the Models and Apps teams, we push it to the W&B Models Registry as a reference artifact. The code below retrieves the weave object version and name for the updated RagModel and uses them to create reference links. A new artifact is then created in W&B with metadata containing the model’s Weave URL. This artifact is logged to the W&B Registry and linked to a designated registry path. Before running the code, ensure the ENTITY and PROJECT variables match your W&B setup, and the target registry path is correctly specified. This process finalizes the workflow by publishing the new RagModel to the W&B ecosystem for easy collaboration and reuse. 
MODELS_OBJECT_VERSION = PUB_REFERENCE.digest  # weave object version
MODELS_OBJECT_NAME = PUB_REFERENCE.name  # weave object name
python
models_url = f"https://wandb.ai/{ENTITY}/{PROJECT}/weave/objects/{MODELS_OBJECT_NAME}/versions/{MODELS_OBJECT_VERSION}"
models_link = (
    f"weave:///{ENTITY}/{PROJECT}/object/{MODELS_OBJECT_NAME}:{MODELS_OBJECT_VERSION}"
)

with wandb.init(project=PROJECT, entity=ENTITY) as run:
    # create new Artifact
    artifact_model = wandb.Artifact(
        name="RagModel",
        type="model",
        description="Models Link from RagModel in Weave",
        metadata={"url": models_url},
    )
    artifact_model.add_reference(models_link, name="model", checksum=False)

    # log new artifact
    run.log_artifact(artifact_model, aliases=[MODELS_OBJECT_VERSION])

    # link to registry
    run.link_artifact(
        artifact_model, target_path="wandb32/wandb-registry-RAG Models/RAG Model"
    )