本指南展示如何使用递归检索遍历节点关系，并基于“引用”获取节点。

节点引用是一个强大的概念。在首次执行检索时，你可能希望检索引用而不是原始文本。你可以让多个引用指向同一个节点。

在本指南中，我们探索节点引用的一些不同用法：

• 分块引用：不同分块大小引用一个更大的分块
• 元数据引用：摘要 + 生成的问题引用一个更大的分块

%pip install llama-index-llms-openai
%pip install llama-index-readers-file

%load_ext autoreload
%autoreload 2
%env OPENAI_API_KEY=YOUR_OPENAI_KEY

如果你在 Colab 中打开这个 Notebook，可能需要先安装 LlamaIndex 🦙。

!pip install llama-index pypdf

加载数据 + 设置

在本节中，我们下载 Llama 2 论文，并创建一组初始节点（分块大小 1024）。

!mkdir -p 'data/'
!wget --user-agent "Mozilla" "https://arxiv.org/pdf/2307.09288.pdf" -O "data/llama2.pdf"

Will not apply HSTS. The HSTS database must be a regular and non-world-writable file.
ERROR: could not open HSTS store at '/home/loganm/.wget-hsts'. HSTS will be disabled.
--2024-01-01 11:13:01--  https://arxiv.org/pdf/2307.09288.pdf
Resolving arxiv.org (arxiv.org)... 151.101.3.42, 151.101.131.42, 151.101.67.42, ...
Connecting to arxiv.org (arxiv.org)|151.101.3.42|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 13661300 (13M) [application/pdf]
Saving to: ‘data/llama2.pdf’

data/llama2.pdf     100%[===================>]  13.03M  27.3MB/s    in 0.5s

2024-01-01 11:13:02 (27.3 MB/s) - ‘data/llama2.pdf’ saved [13661300/13661300]

from pathlib import Path
from llama_index.readers.file import PDFReader
from llama_index.core.response.notebook_utils import display_source_node
from llama_index.core.retrievers import RecursiveRetriever
from llama_index.core.query_engine import RetrieverQueryEngine
from llama_index.core import VectorStoreIndex
from llama_index.llms.openai import OpenAI
import json

loader = PDFReader()
docs0 = loader.load_data(file=Path("./data/llama2.pdf"))

from llama_index.core import Document

doc_text = "\n\n".join([d.get_content() for d in docs0])
docs = [Document(text=doc_text)]

from llama_index.core.node_parser import SentenceSplitter
from llama_index.core.schema import IndexNode

node_parser = SentenceSplitter(chunk_size=1024)

base_nodes = node_parser.get_nodes_from_documents(docs)
# set node ids to be a constant
for idx, node in enumerate(base_nodes):
    node.id_ = f"node-{idx}"

from llama_index.core.embeddings import resolve_embed_model

embed_model = resolve_embed_model("local:BAAI/bge-small-en")
llm = OpenAI(model="gpt-3.5-turbo")

基线检索器

定义一个基线检索器，它仅按嵌入相似度获取前 k 个原始文本节点。

base_index = VectorStoreIndex(base_nodes, embed_model=embed_model)
base_retriever = base_index.as_retriever(similarity_top_k=2)

retrievals = base_retriever.retrieve(
    "Can you tell me about the key concepts for safety finetuning"
)

for n in retrievals:
    display_source_node(n, source_length=1500)

Node ID: node-26
Similarity: 0.8581930837671874
Text: AsLLMsareintegratedanddeployed,welookforwardto continuing research that will amplify their potential for positive impact on these important social issues. 4.2 Safety Fine-Tuning In this section, we describe our approach to safety fine-tuning, including safety categories, annotation guidelines,andthetechniquesweusetomitigatesafetyrisks. Weemployaprocesssimilartothegeneral fine-tuning methods as described in Section 3, with some notable differences related to safety concerns. Specifically, we use the following techniques in safety fine-tuning: 1.Supervised Safety Fine-Tuning : We initialize by gathering adversarial prompts and safe demonstra- tions that are then included in the general supervised fine-tuning process (Section 3.1). This teaches themodeltoalignwithoursafetyguidelinesevenbeforeRLHF,andthuslaysthefoundationfor high-quality human preference data annotation. 2.Safety RLHF : Subsequently, we integrate safety in the general RLHF pipeline described in Sec- tion 3.2.2. This includes training a safety-specific reward model and gathering more challenging adversarial prompts for rejection sampling style fine-tuning and PPO optimization. 3.SafetyContextDistillation : Finally,werefineourRLHFpipelinewithcontextdistillation(Askell etal.,2021b). Thisinvolvesgeneratingsafermodelresponsesbyprefixingapromptwithasafety preprompt, e.g., “You are a safe and responsible assistant,” and then fine-tuning the model on the safer responses without the preprompt, which essentially distill…

Node ID: node-33
Similarity: 0.8442806881920294
Text: Examples of evaluation prompts and responses can be found in Appendix A.4.6. We then asked raters to judge models for safety violations on a five-point Likert scale with the following definitions: •5- No safety violations and very helpful •4- No safety violations and only minor non-safety issues •3- No safety violations but not helpful or other major non-safety issues •2- Mild or moderate safety violations 29

(a)Overall violation percentage. (b)Overall safety and helpfulness mean rating. Figure17: Overallsafetymeasures. Left: Llama 2-Chat haslowviolationpercentageoverallacrossmodel sizes.Right: Llama 2-Chat has high safety and helpfulness mean rating overall across model sizes. It is important to note that these results are subject to limitations of the prompt set, subjectivity of the review guidelines, and subjectivity of individual raters. •1- Severe safety violations We consider a rating of 1 or 2 as violation and use violation percentage as our main evaluation metric, with themeanratingasasupplement. Eachexampleisannotatedbythreeannotatorsandwetakethemajority votetodetermineiftheresponseisviolatingornot. WeusedGwet’sAC1/2statistictomeasureinter-rater reliability(IRR)asinthehelpfulnesshumanevaluation. TheIRRscoresrangefrom 0.70to0.95depending on the annotation batch, indicating a high degree of agreement among annotators on safety assessments. OnLlama 2-Chat annotations, the average IRR is 0.92according to Gwet’s AC2 measure. We see lower IRR scoresonbatcheswherethemo…

query_engine_base = RetrieverQueryEngine.from_args(base_retriever, llm=llm)

response = query_engine_base.query(
    "Can you tell me about the key concepts for safety finetuning"
)
print(str(response))

The key concepts for safety fine-tuning include supervised safety fine-tuning, safety RLHF (Reinforcement Learning from Human Feedback), and safety context distillation. In supervised safety fine-tuning, adversarial prompts and safe demonstrations are gathered and included in the general supervised fine-tuning process. This helps the model align with safety guidelines and lays the foundation for high-quality human preference data annotation. Safety RLHF involves integrating safety in the general RLHF pipeline, which includes training a safety-specific reward model and gathering more challenging adversarial prompts for rejection sampling style fine-tuning and PPO (Proximal Policy Optimization) optimization. Safety context distillation is the final step, where the RLHF pipeline is refined with context distillation. This involves generating safer model responses by prefixing a prompt with a safety preprompt and then fine-tuning the model on the safer responses without the preprompt.

分块引用：较小的子分块引用更大的父分块

在此用法示例中，我们展示如何构建一个由较小分块指向较大父分块的图。

在查询时，我们检索较小的分块，但会跟随引用找到较大的分块。这使我们能够获得更多上下文以进行综合。

sub_chunk_sizes = [128, 256, 512]
sub_node_parsers = [
    SentenceSplitter(chunk_size=c, chunk_overlap=20) for c in sub_chunk_sizes
]

all_nodes = []
for base_node in base_nodes:
    for n in sub_node_parsers:
        sub_nodes = n.get_nodes_from_documents([base_node])
        sub_inodes = [
            IndexNode.from_text_node(sn, base_node.node_id) for sn in sub_nodes
        ]
        all_nodes.extend(sub_inodes)

    # also add original node to node
    original_node = IndexNode.from_text_node(base_node, base_node.node_id)
    all_nodes.append(original_node)

all_nodes_dict = {n.node_id: n for n in all_nodes}

vector_index_chunk = VectorStoreIndex(all_nodes, embed_model=embed_model)

vector_retriever_chunk = vector_index_chunk.as_retriever(similarity_top_k=2)

retriever_chunk = RecursiveRetriever(
    "vector",
    retriever_dict={"vector": vector_retriever_chunk},
    node_dict=all_nodes_dict,
    verbose=True,
)

nodes = retriever_chunk.retrieve(
    "Can you tell me about the key concepts for safety finetuning"
)
for node in nodes:
    display_source_node(node, source_length=2000)

Retrieving with query id None: Can you tell me about the key concepts for safety finetuning
Retrieved node with id, entering: node-26
Retrieving with query id node-26: Can you tell me about the key concepts for safety finetuning
Retrieved node with id, entering: node-1
Retrieving with query id node-1: Can you tell me about the key concepts for safety finetuning


Node ID: node-26
Similarity: 0.8809071991986446
Text: AsLLMsareintegratedanddeployed,welookforwardto continuing research that will amplify their potential for positive impact on these important social issues. 4.2 Safety Fine-Tuning In this section, we describe our approach to safety fine-tuning, including safety categories, annotation guidelines,andthetechniquesweusetomitigatesafetyrisks. Weemployaprocesssimilartothegeneral fine-tuning methods as described in Section 3, with some notable differences related to safety concerns. Specifically, we use the following techniques in safety fine-tuning: 1.Supervised Safety Fine-Tuning : We initialize by gathering adversarial prompts and safe demonstra- tions that are then included in the general supervised fine-tuning process (Section 3.1). This teaches themodeltoalignwithoursafetyguidelinesevenbeforeRLHF,andthuslaysthefoundationfor high-quality human preference data annotation. 2.Safety RLHF : Subsequently, we integrate safety in the general RLHF pipeline described in Sec- tion 3.2.2. This includes training a safety-specific reward model and gathering more challenging adversarial prompts for rejection sampling style fine-tuning and PPO optimization. 3.SafetyContextDistillation : Finally,werefineourRLHFpipelinewithcontextdistillation(Askell etal.,2021b). Thisinvolvesgeneratingsafermodelresponsesbyprefixingapromptwithasafety preprompt, e.g., “You are a safe and responsible assistant,” and then fine-tuning the model on the safer responses without the preprompt, which essentially distillsthe safety preprompt (context) into the model. Weuseatargetedapproachthatallowsoursafetyrewardmodeltochoosewhethertouse context distillation for each sample. 4.2.1 Safety Categories and Annotation Guidelines Based on limitations of LLMs known from prior work, we design instructions for our annotation team to createadversarialpromptsalongtwodimensions: a riskcategory ,orpotentialtopicaboutwhichtheLLM couldproduceunsafecontent;andan attackvector ,orquestionstyletocoverdifferentvarietiesofprompts …

Node ID: node-1
Similarity: 0.8744334039911964
Text: … … … … … … … … … … … … 9 3.2 Reinforcement Learning with Human Feedback (RLHF) … … … … … … … 9 3.3 System Message for Multi-Turn Consistency … … … … … … … … … . . 16 3.4 RLHF Results … … … … … … … … … … … … … … … . 17 4 Safety 20 4.1 Safety in Pretraining … … … … … … … … … … … … … … 20 4.2 Safety Fine-Tuning … … … … … … … … … … … … … … . 23 4.3 Red Teaming … … … … … … … … … … … … … … … . . 28 4.4 Safety Evaluation of Llama 2-Chat … … … … … … … … … … … . 29 5 Discussion 32 5.1 Learnings and Observations … … … … … … … … … … … … . . 32 5.2 Limitations and Ethical Considerations … … … … … … … … … … . 34 5.3 Responsible Release Strategy … … … … … … … … … … … … . 35 6 Related Work 35 7 Conclusion 36 A Appendix 46 A.1 Contributions … … … … … … … … … … … … …

query_engine_chunk = RetrieverQueryEngine.from_args(retriever_chunk, llm=llm)

response = query_engine_chunk.query(
    "Can you tell me about the key concepts for safety finetuning"
)
print(str(response))

Retrieving with query id None: Can you tell me about the key concepts for safety finetuning
Retrieved node with id, entering: node-26
Retrieving with query id node-26: Can you tell me about the key concepts for safety finetuning
Retrieved node with id, entering: node-1
Retrieving with query id node-1: Can you tell me about the key concepts for safety finetuning
The key concepts for safety fine-tuning include supervised safety fine-tuning, safety RLHF (Reinforcement Learning with Human Feedback), and safety context distillation. Supervised safety fine-tuning involves gathering adversarial prompts and safe demonstrations to teach the model to align with safety guidelines. Safety RLHF integrates safety into the general RLHF pipeline by training a safety-specific reward model and gathering challenging adversarial prompts for rejection sampling style fine-tuning and PPO optimization. Safety context distillation involves generating safer model responses by prefixing a prompt with a safety preprompt and fine-tuning the model on the safer responses without the preprompt. These techniques aim to mitigate safety risks and improve the model's ability to provide safe and responsible responses.

元数据引用：摘要 + 生成的问题引用更大的分块

在此用法示例中，我们展示如何定义引用源节点的额外上下文。

这些额外上下文包括摘要以及生成的问题。

在查询时，我们检索较小的分块，但会跟随引用找到较大的分块。这使我们能够获得更多上下文以进行综合。

import nest_asyncio

nest_asyncio.apply()

from llama_index.core.node_parser import SentenceSplitter
from llama_index.core.schema import IndexNode
from llama_index.core.extractors import (
    SummaryExtractor,
    QuestionsAnsweredExtractor,
)

extractors = [
    SummaryExtractor(summaries=["self"], show_progress=True),
    QuestionsAnsweredExtractor(questions=5, show_progress=True),
]

# run metadata extractor across base nodes, get back dictionaries
node_to_metadata = {}
for extractor in extractors:
    metadata_dicts = extractor.extract(base_nodes)
    for node, metadata in zip(base_nodes, metadata_dicts):
        if node.node_id not in node_to_metadata:
            node_to_metadata[node.node_id] = metadata
        else:
            node_to_metadata[node.node_id].update(metadata)

100%|██████████| 93/93 [01:13<00:00,  1.27it/s]
100%|██████████| 93/93 [00:49<00:00,  1.88it/s]

# cache metadata dicts
def save_metadata_dicts(path, data):
    with open(path, "w") as fp:
        json.dump(data, fp)

def load_metadata_dicts(path):
    with open(path, "r") as fp:
        data = json.load(fp)
    return data

save_metadata_dicts("data/llama2_metadata_dicts.json", node_to_metadata)

metadata_dicts = load_metadata_dicts("data/llama2_metadata_dicts.json")

# all nodes consists of source nodes, along with metadata
import copy

all_nodes = copy.deepcopy(base_nodes)
for node_id, metadata in node_to_metadata.items():
    for val in metadata.values():
        all_nodes.append(IndexNode(text=val, index_id=node_id))

all_nodes_dict = {n.node_id: n for n in all_nodes}

## Load index into vector index
from llama_index.core import VectorStoreIndex
from llama_index.llms.openai import OpenAI

llm = OpenAI(model="gpt-3.5-turbo")

vector_index_metadata = VectorStoreIndex(all_nodes)

vector_retriever_metadata = vector_index_metadata.as_retriever(
    similarity_top_k=2
)

retriever_metadata = RecursiveRetriever(
    "vector",
    retriever_dict={"vector": vector_retriever_metadata},
    node_dict=all_nodes_dict,
    verbose=False,
)

nodes = retriever_metadata.retrieve(
    "Can you tell me about the key concepts for safety finetuning"
)
for node in nodes:
    display_source_node(node, source_length=2000)

Node ID: node-26
Similarity: 0.8727061238826861
Text: AsLLMsareintegratedanddeployed,welookforwardto continuing research that will amplify their potential for positive impact on these important social issues. 4.2 Safety Fine-Tuning In this section, we describe our approach to safety fine-tuning, including safety categories, annotation guidelines,andthetechniquesweusetomitigatesafetyrisks. Weemployaprocesssimilartothegeneral fine-tuning methods as described in Section 3, with some notable differences related to safety concerns. Specifically, we use the following techniques in safety fine-tuning: 1.Supervised Safety Fine-Tuning : We initialize by gathering adversarial prompts and safe demonstra- tions that are then included in the general supervised fine-tuning process (Section 3.1). This teaches themodeltoalignwithoursafetyguidelinesevenbeforeRLHF,andthuslaysthefoundationfor high-quality human preference data annotation. 2.Safety RLHF : Subsequently, we integrate safety in the general RLHF pipeline described in Sec- tion 3.2.2. This includes training a safety-specific reward model and gathering more challenging adversarial prompts for rejection sampling style fine-tuning and PPO optimization. 3.SafetyContextDistillation : Finally,werefineourRLHFpipelinewithcontextdistillation(Askell etal.,2021b). Thisinvolvesgeneratingsafermodelresponsesbyprefixingapromptwithasafety preprompt, e.g., “You are a safe and responsible assistant,” and then fine-tuning the model on the safer responses without the preprompt, which essentially distillsthe safety preprompt (context) into the model. Weuseatargetedapproachthatallowsoursafetyrewardmodeltochoosewhethertouse context distillation for each sample. 4.2.1 Safety Categories and Annotation Guidelines Based on limitations of LLMs known from prior work, we design instructions for our annotation team to createadversarialpromptsalongtwodimensions: a riskcategory ,orpotentialtopicaboutwhichtheLLM couldproduceunsafecontent;andan attackvector ,orquestionstyletocoverdifferentvarietiesofprompts …

Node ID: node-26
Similarity: 0.8586079224453517
Text: AsLLMsareintegratedanddeployed,welookforwardto continuing research that will amplify their potential for positive impact on these important social issues. 4.2 Safety Fine-Tuning In this section, we describe our approach to safety fine-tuning, including safety categories, annotation guidelines,andthetechniquesweusetomitigatesafetyrisks. Weemployaprocesssimilartothegeneral fine-tuning methods as described in Section 3, with some notable differences related to safety concerns. Specifically, we use the following techniques in safety fine-tuning: 1.Supervised Safety Fine-Tuning : We initialize by gathering adversarial prompts and safe demonstra- tions that are then included in the general supervised fine-tuning process (Section 3.1). This teaches themodeltoalignwithoursafetyguidelinesevenbeforeRLHF,andthuslaysthefoundationfor high-quality human preference data annotation. 2.Safety RLHF : Subsequently, we integrate safety in the general RLHF pipeline described in Sec- tion 3.2.2. This includes training a safety-specific reward model and gathering more challenging adversarial prompts for rejection sampling style fine-tuning and PPO optimization. 3.SafetyContextDistillation : Finally,werefineourRLHFpipelinewithcontextdistillation(Askell etal.,2021b). Thisinvolvesgeneratingsafermodelresponsesbyprefixingapromptwithasafety preprompt, e.g., “You are a safe and responsible assistant,” and then fine-tuning the model on the safer responses without the preprompt, which essentially distillsthe safety preprompt (context) into the model. Weuseatargetedapproachthatallowsoursafetyrewardmodeltochoosewhethertouse context distillation for each sample. 4.2.1 Safety Categories and Annotation Guidelines Based on limitations of LLMs known from prior work, we design instructions for our annotation team to createadversarialpromptsalongtwodimensions: a riskcategory ,orpotentialtopicaboutwhichtheLLM couldproduceunsafecontent;andan attackvector ,orquestionstyletocoverdifferentvarietiesofprompts …

query_engine_metadata = RetrieverQueryEngine.from_args(
    retriever_metadata, llm=llm
)

response = query_engine_metadata.query(
    "Can you tell me about the key concepts for safety finetuning"
)
print(str(response))

The key concepts for safety fine-tuning include supervised safety fine-tuning, safety RLHF (Reinforcement Learning from Human Feedback), and safety context distillation. Supervised safety fine-tuning involves gathering adversarial prompts and safe demonstrations to train the model to align with safety guidelines. Safety RLHF integrates safety into the RLHF pipeline by training a safety-specific reward model and gathering challenging adversarial prompts for fine-tuning and optimization. Safety context distillation involves generating safer model responses by prefixing a prompt with a safety preprompt and fine-tuning the model on the safer responses without the preprompt. These concepts are used to mitigate safety risks and improve the model's ability to produce safe and helpful responses.

评估

我们评估递归检索 + 节点引用方法的效果。我们同时评估分块引用和元数据引用。我们使用嵌入相似度查找来检索引用节点。

我们将两种方法与直接获取原始节点的基线检索器进行比较。

在指标方面，我们使用命中率和 MRR 进行评估。

数据集生成

我们首先从文本分块集合中生成一个问答数据集。

from llama_index.core.evaluation import (
    generate_question_context_pairs,
    EmbeddingQAFinetuneDataset,
)
from llama_index.llms.openai import OpenAI

import nest_asyncio

nest_asyncio.apply()

eval_dataset = generate_question_context_pairs(
    base_nodes, OpenAI(model="gpt-3.5-turbo")
)

100%|██████████| 93/93 [02:08<00:00,  1.38s/it]

eval_dataset.save_json("data/llama2_eval_dataset.json")

# optional
eval_dataset = EmbeddingQAFinetuneDataset.from_json(
    "data/llama2_eval_dataset.json"
)

比较结果

我们对每个检索器运行评估，以衡量命中率和 MRR。

我们发现，使用节点引用（分块引用或元数据引用）的检索器往往比直接检索原始分块表现更好。

import pandas as pd
from llama_index.core.evaluation import (
    RetrieverEvaluator,
    get_retrieval_results_df,
)

# set vector retriever similarity top k to higher
top_k = 10

def display_results(names, results_arr):
    """Display results from evaluate."""

    hit_rates = []
    mrrs = []
    for name, eval_results in zip(names, results_arr):
        metric_dicts = []
        for eval_result in eval_results:
            metric_dict = eval_result.metric_vals_dict
            metric_dicts.append(metric_dict)
        results_df = pd.DataFrame(metric_dicts)

        hit_rate = results_df["hit_rate"].mean()
        mrr = results_df["mrr"].mean()
        hit_rates.append(hit_rate)
        mrrs.append(mrr)

    final_df = pd.DataFrame(
        {"retrievers": names, "hit_rate": hit_rates, "mrr": mrrs}
    )
    display(final_df)

vector_retriever_chunk = vector_index_chunk.as_retriever(
    similarity_top_k=top_k
)
retriever_chunk = RecursiveRetriever(
    "vector",
    retriever_dict={"vector": vector_retriever_chunk},
    node_dict=all_nodes_dict,
    verbose=True,
)
retriever_evaluator = RetrieverEvaluator.from_metric_names(
    ["mrr", "hit_rate"], retriever=retriever_chunk
)
# try it out on an entire dataset
results_chunk = await retriever_evaluator.aevaluate_dataset(
    eval_dataset, show_progress=True
)

vector_retriever_metadata = vector_index_metadata.as_retriever(
    similarity_top_k=top_k
)
retriever_metadata = RecursiveRetriever(
    "vector",
    retriever_dict={"vector": vector_retriever_metadata},
    node_dict=all_nodes_dict,
    verbose=True,
)
retriever_evaluator = RetrieverEvaluator.from_metric_names(
    ["mrr", "hit_rate"], retriever=retriever_metadata
)
# try it out on an entire dataset
results_metadata = await retriever_evaluator.aevaluate_dataset(
    eval_dataset, show_progress=True
)

base_retriever = base_index.as_retriever(similarity_top_k=top_k)
retriever_evaluator = RetrieverEvaluator.from_metric_names(
    ["mrr", "hit_rate"], retriever=base_retriever
)
# try it out on an entire dataset
results_base = await retriever_evaluator.aevaluate_dataset(
    eval_dataset, show_progress=True
)

100%|██████████| 194/194 [00:09<00:00, 19.86it/s]

full_results_df = get_retrieval_results_df(
    [
        "Base Retriever",
        "Retriever (Chunk References)",
        "Retriever (Metadata References)",
    ],
    [results_base, results_chunk, results_metadata],
)
display(full_results_df)

.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}

	retrievers	hit_rate	mrr
0	Base Retriever	0.778351	0.563103
1	Retriever (Chunk References)	0.896907	0.691114
2	Retriever (Metadata References)	0.891753	0.718440