本 notebook 展示如何在 Llamaindex 中使用 Predibase 托管的 LLM。你可以将 Predibase 添加到现有的 Llamaindex 工作流，以便：

1. 无需繁琐操作即可部署和查询预训练或自定义开源 LLM
2. 将端到端的检索增强生成（RAG）系统投入运营
3. 只需几行代码即可微调你自己的 LLM

开始使用

1. 在此注册免费的 Predibase 账户
2. 创建账户
3. 前往 Settings > My profile 并生成新的 API Token。

%pip install llama-index-llms-predibase

!pip install llama-index --quiet
!pip install predibase --quiet
!pip install sentence-transformers --quiet

import os

os.environ["PREDIBASE_API_TOKEN"] = "{PREDIBASE_API_TOKEN}"
from llama_index.llms.predibase import PredibaseLLM

流程 1：直接查询 Predibase LLM

# Predibase-hosted fine-tuned adapter example
llm = PredibaseLLM(
    model_name="mistral-7b",
    predibase_sdk_version=None,  # optional parameter (defaults to the latest Predibase SDK version if omitted)
    adapter_id="e2e_nlg",  # adapter_id is optional
    adapter_version=1,  # optional parameter (applies to Predibase only)
    api_token=None,  # optional parameter for accessing services hosting adapters (e.g., HuggingFace)
    max_new_tokens=512,
    temperature=0.3,
)
# The `model_name` parameter is the Predibase "serverless" base_model ID
# (see https://docs.predibase.com/user-guide/inference/models for the catalog).
# You can also optionally specify a fine-tuned adapter that's hosted on Predibase or HuggingFace
# In the case of Predibase-hosted adapters, you must also specify the adapter_version

# HuggingFace-hosted fine-tuned adapter example
llm = PredibaseLLM(
    model_name="mistral-7b",
    predibase_sdk_version=None,  # optional parameter (defaults to the latest Predibase SDK version if omitted)
    adapter_id="predibase/e2e_nlg",  # adapter_id is optional
    api_token=os.environ.get(
        "HUGGING_FACE_HUB_TOKEN"
    ),  # optional parameter for accessing services hosting adapters (e.g., HuggingFace)
    max_new_tokens=512,
    temperature=0.3,
)
# The `model_name` parameter is the Predibase "serverless" base_model ID
# (see https://docs.predibase.com/user-guide/inference/models for the catalog).
# You can also optionally specify a fine-tuned adapter that's hosted on Predibase or HuggingFace
# In the case of Predibase-hosted adapters, you can also specify the adapter_version (assumed latest if omitted)

result = llm.complete("Can you recommend me a nice dry white wine?")
print(result)

流程 2：使用 Predibase LLM 进行检索增强生成（RAG）

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
from llama_index.core.embeddings import resolve_embed_model
from llama_index.core.node_parser import SentenceSplitter

下载数据

!mkdir -p 'data/paul_graham/'
!wget 'https://raw.githubusercontent.com/run-llama/llama_index/main/docs/examples/data/paul_graham/paul_graham_essay.txt' -O 'data/paul_graham/paul_graham_essay.txt'

加载文档

documents = SimpleDirectoryReader("./data/paul_graham/").load_data()

配置 Predibase LLM

# Predibase-hosted fine-tuned adapter
llm = PredibaseLLM(
    model_name="mistral-7b",
    predibase_sdk_version=None,  # optional parameter (defaults to the latest Predibase SDK version if omitted)
    adapter_id="e2e_nlg",  # adapter_id is optional
    api_token=None,  # optional parameter for accessing services hosting adapters (e.g., HuggingFace)
    temperature=0.3,
    context_window=1024,
)

# HuggingFace-hosted fine-tuned adapter
llm = PredibaseLLM(
    model_name="mistral-7b",
    predibase_sdk_version=None,  # optional parameter (defaults to the latest Predibase SDK version if omitted)
    adapter_id="predibase/e2e_nlg",  # adapter_id is optional
    api_token=os.environ.get(
        "HUGGING_FACE_HUB_TOKEN"
    ),  # optional parameter for accessing services hosting adapters (e.g., HuggingFace)
    temperature=0.3,
    context_window=1024,
)

embed_model = resolve_embed_model("local:BAAI/bge-small-en-v1.5")
splitter = SentenceSplitter(chunk_size=1024)

设置并查询索引

index = VectorStoreIndex.from_documents(
    documents, transformations=[splitter], embed_model=embed_model
)
query_engine = index.as_query_engine(llm=llm)
response = query_engine.query("What did the author do growing up?")

print(response)