Router Retriever

In this guide, we define a custom router retriever that selects one or more candidate retrievers in order to execute a given query.

The router (BaseSelector) module uses the LLM to dynamically make decisions on which underlying retrieval tools to use. This can be helpful to select one out of a diverse range of data sources. This can also be helpful to aggregate retrieval results across a variety of data sources (if a multi-selector module is used).

This notebook is very similar to the RouterQueryEngine notebook.

Setup

If you’re opening this Notebook on colab, you will probably need to install LlamaIndex 🦙.

%pip install llama-index-llms-openai

!pip install llama-index

# NOTE: This is ONLY necessary in jupyter notebook.
# Details: Jupyter runs an event-loop behind the scenes.
#          This results in nested event-loops when we start an event-loop to make async queries.
#          This is normally not allowed, we use nest_asyncio to allow it for convenience.
import nest_asyncio

nest_asyncio.apply()

import logging
import sys

logging.basicConfig(stream=sys.stdout, level=logging.INFO)
logging.getLogger().handlers = []
logging.getLogger().addHandler(logging.StreamHandler(stream=sys.stdout))

from llama_index.core import (
    VectorStoreIndex,
    SimpleDirectoryReader,
    StorageContext,
    SimpleKeywordTableIndex,
)
from llama_index.core import SummaryIndex
from llama_index.core.node_parser import SentenceSplitter
from llama_index.llms.openai import OpenAI

Note: NumExpr detected 12 cores but "NUMEXPR_MAX_THREADS" not set, so enforcing safe limit of 8.
NumExpr defaulting to 8 threads.

Download Data

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

Load Data

We first show how to convert a Document into a set of Nodes, and insert into a DocumentStore.

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

# initialize LLM + splitter
llm = OpenAI(model="gpt-4")
splitter = SentenceSplitter(chunk_size=1024)
nodes = splitter.get_nodes_from_documents(documents)

# initialize storage context (by default it's in-memory)
storage_context = StorageContext.from_defaults()
storage_context.docstore.add_documents(nodes)

# define
summary_index = SummaryIndex(nodes, storage_context=storage_context)
vector_index = VectorStoreIndex(nodes, storage_context=storage_context)
keyword_index = SimpleKeywordTableIndex(nodes, storage_context=storage_context)

list_retriever = summary_index.as_retriever()
vector_retriever = vector_index.as_retriever()
keyword_retriever = keyword_index.as_retriever()

from llama_index.core.tools import RetrieverTool

list_tool = RetrieverTool.from_defaults(
    retriever=list_retriever,
    description=(
        "Will retrieve all context from Paul Graham's essay on What I Worked"
        " On. Don't use if the question only requires more specific context."
    ),
)
vector_tool = RetrieverTool.from_defaults(
    retriever=vector_retriever,
    description=(
        "Useful for retrieving specific context from Paul Graham essay on What"
        " I Worked On."
    ),
)
keyword_tool = RetrieverTool.from_defaults(
    retriever=keyword_retriever,
    description=(
        "Useful for retrieving specific context from Paul Graham essay on What"
        " I Worked On (using entities mentioned in query)"
    ),
)

Define Selector Module for Routing

There are several selectors available, each with some distinct attributes.

The LLM selectors use the LLM to output a JSON that is parsed, and the corresponding indexes are queried.

The Pydantic selectors (currently only supported by gpt-4-0613 and gpt-3.5-turbo-0613 (the default)) use the OpenAI Function Call API to produce pydantic selection objects, rather than parsing raw JSON.

Here we use PydanticSingleSelector/PydanticMultiSelector but you can use the LLM-equivalents as well.

from llama_index.core.selectors import LLMSingleSelector, LLMMultiSelector
from llama_index.core.selectors import (
    PydanticMultiSelector,
    PydanticSingleSelector,
)
from llama_index.core.retrievers import RouterRetriever
from llama_index.core.response.notebook_utils import display_source_node

PydanticSingleSelector

retriever = RouterRetriever(
    selector=PydanticSingleSelector.from_defaults(llm=llm),
    retriever_tools=[
        list_tool,
        vector_tool,
    ],
)

# will retrieve all context from the author's life
nodes = retriever.retrieve(
    "Can you give me all the context regarding the author's life?"
)
for node in nodes:
    display_source_node(node)

Selecting retriever 0: This choice is most relevant as it mentions retrieving all context from the essay, which could include information about the author's life..

Node ID: 7d07d325-489e-4157-a745-270e2066a643
Similarity: None
Text: What I Worked On

February 2021

Before college the two main things I worked on, outside of schoo…

Node ID: 01f0900b-db83-450b-a088-0473f16882d7
Similarity: None
Text: showed Terry Winograd using SHRDLU. I haven’t tried rereading The Moon is a Harsh Mistress, so I …

Node ID: b2549a68-5fef-4179-b027-620ebfa6e346
Similarity: None
Text: Science is an uneasy alliance between two halves, theory and systems. The theory people prove thi…

Node ID: 4f1e9f0d-9bc6-4169-b3b6-4f169bbfa391
Similarity: None
Text: been explored. But all I wanted was to get out of grad school, and my rapidly written dissertatio…

Node ID: e20c99f9-5e80-4c92-8cc0-03d2a527131e
Similarity: None
Text: stop there, of course, or you get merely photographic accuracy, and what makes a still life inter…

Node ID: dbdf341a-f340-49f9-961f-16b9a51eea2d
Similarity: None
Text: that big, bureaucratic customers are a dangerous source of money, and that there’s not much overl…

Node ID: ed341d3a-9dda-49c1-8611-0ab40d04f08a
Similarity: None
Text: about money, because I could sense that Interleaf was on the way down. Freelance Lisp hacking wor…

Node ID: d69e02d3-2732-4567-a360-893c14ae157b
Similarity: None
Text: a web app, is common now, but at the time it wasn’t clear that it was even possible. To find out,…

Node ID: df9e00a5-e795-40a1-9a6b-8184d1b1e7c0
Similarity: None
Text: have to integrate with any other software except Robert’s and Trevor’s, so it was quite fun to wo…

Node ID: 38f2699b-0878-499b-90ee-821cb77e387b
Similarity: None
Text: all too keenly aware of the near-death experiences we seemed to have every few months. Nor had I …

Node ID: be04d6a9-1fc7-4209-9df2-9c17a453699a
Similarity: None
Text: for a second still life, painted from the same objects (which hopefully hadn’t rotted yet).

Mean…

Node ID: 42344911-8a7c-4e9b-81a8-0fcf40ab7690
Similarity: None
Text: which I’d created years before using Viaweb but had never used for anything. In one day it got 30…

Node ID: 9ec3df49-abf9-47f4-b0c2-16687882742a
Similarity: None
Text: I didn’t know but would turn out to like a lot: a woman called Jessica Livingston. A couple days …

Node ID: d0cf6975-5261-4fb2-aae3-f3230090fb64
Similarity: None
Text: of readers, but professional investors are thinking “Wow, that means they got all the returns.” B…

Node ID: 607d0480-7eee-4fb4-965d-3cb585fda62c
Similarity: None
Text: to the “YC GDP,” but as YC grows this becomes less and less of a joke. Now lots of startups get t…

Node ID: 730a49c9-55f7-4416-ab91-1d0c96e704c8
Similarity: None
Text: So this set me thinking. It was true that on my current trajectory, YC would be the last thing I …

Node ID: edbe8c67-e373-42bf-af98-276b559cc08b
Similarity: None
Text: operators you need? The Lisp that John McCarthy invented, or more accurately discovered, is an an…

Node ID: 175a4375-35ec-45a0-a90c-15611505096b
Similarity: None
Text: Like McCarthy’s original Lisp, it’s a spec rather than an implementation, although like McCarthy’…

Node ID: 0cb367f9-0aac-422b-9243-0eaa7be15090
Similarity: None
Text: must tell readers things they don’t already know, and some people dislike being told such things…

Node ID: 67afd4f1-9fa1-4e76-87ac-23b115823e6c
Similarity: None
Text: 1960 paper.

But if so there’s no reason to suppose that this is the limit of the language that m…

nodes = retriever.retrieve("What did Paul Graham do after RISD?")
for node in nodes:
    display_source_node(node)

Selecting retriever 1: The question asks for a specific detail from Paul Graham's essay on 'What I Worked On'. Therefore, the second choice, which is useful for retrieving specific context, is the most relevant..

Node ID: 22d20835-7de6-4cf7-92de-2bee339f3157
Similarity: 0.8017176790752668
Text: that big, bureaucratic customers are a dangerous source of money, and that there’s not much overl…

Node ID: bf818c58-5d5b-4458-acbc-d87cc67a36ca
Similarity: 0.7935885352785799
Text: So this set me thinking. It was true that on my current trajectory, YC would be the last thing I …

PydanticMultiSelector

retriever = RouterRetriever(
    selector=PydanticMultiSelector.from_defaults(llm=llm),
    retriever_tools=[list_tool, vector_tool, keyword_tool],
)

nodes = retriever.retrieve(
    "What were noteable events from the authors time at Interleaf and YC?"
)
for node in nodes:
    display_source_node(node)

Selecting retriever 1: This choice is relevant as it allows for retrieving specific context from the essay, which is needed to answer the question about notable events at Interleaf and YC..
Selecting retriever 2: This choice is also relevant as it allows for retrieving specific context using entities mentioned in the query, which in this case are 'Interleaf' and 'YC'..
> Starting query: What were noteable events from the authors time at Interleaf and YC?
query keywords: ['interleaf', 'events', 'noteable', 'yc']
> Extracted keywords: ['interleaf', 'yc']

Node ID: fbdd25ed-1ecb-4528-88da-34f581c30782
Similarity: None
Text: So this set me thinking. It was true that on my current trajectory, YC would be the last thing I …

Node ID: 4ce91b17-131f-4155-b7b5-8917cdc612b1
Similarity: None
Text: to the “YC GDP,” but as YC grows this becomes less and less of a joke. Now lots of startups get t…

Node ID: 9fe6c152-28d4-4006-8a1a-43bb72655438
Similarity: None
Text: stop there, of course, or you get merely photographic accuracy, and what makes a still life inter…

Node ID: d11cd2e2-1dd2-4c3b-863f-246fe3856f49
Similarity: None
Text: of readers, but professional investors are thinking “Wow, that means they got all the returns.” B…

Node ID: 2bfbab04-cb71-4641-9bd9-52c75b3a9250
Similarity: None
Text: must tell readers things they don’t already know, and some people dislike being told such things…

nodes = retriever.retrieve(
    "What were noteable events from the authors time at Interleaf and YC?"
)
for node in nodes:
    display_source_node(node)

Selecting retriever 1: This choice is relevant as it allows for retrieving specific context from the essay, which is needed to answer the question about notable events at Interleaf and YC..
Selecting retriever 2: This choice is also relevant as it allows for retrieving specific context using entities mentioned in the query, which in this case are 'Interleaf' and 'YC'..
> Starting query: What were noteable events from the authors time at Interleaf and YC?
query keywords: ['interleaf', 'yc', 'events', 'noteable']
> Extracted keywords: ['interleaf', 'yc']

Node ID: 49882a2c-bb95-4ff3-9df1-2a40ddaea408
Similarity: None
Text: So this set me thinking. It was true that on my current trajectory, YC would be the last thing I …

Node ID: d11aced1-e630-4109-8ec8-194e975b9851
Similarity: None
Text: to the “YC GDP,” but as YC grows this becomes less and less of a joke. Now lots of startups get t…

Node ID: 8aa6cc91-8e9c-4470-b6d5-4360ed13fefd
Similarity: None
Text: stop there, of course, or you get merely photographic accuracy, and what makes a still life inter…

Node ID: e37465de-c79a-4714-a402-fbd5f52800a2
Similarity: None
Text: must tell readers things they don’t already know, and some people dislike being told such things…

Node ID: e0ac7fb6-84fc-4763-bca6-b68f300ec7b7
Similarity: None
Text: of readers, but professional investors are thinking “Wow, that means they got all the returns.” B…

nodes = await retriever.aretrieve(
    "What were noteable events from the authors time at Interleaf and YC?"
)
for node in nodes:
    display_source_node(node)

Selecting retriever 1: This choice is relevant as it allows for retrieving specific context from the essay, which is needed to answer the question about notable events at Interleaf and YC..
Selecting retriever 2: This choice is also relevant as it allows for retrieving specific context using entities mentioned in the query, which in this case are 'Interleaf' and 'YC'..
> Starting query: What were noteable events from the authors time at Interleaf and YC?
query keywords: ['events', 'interleaf', 'yc', 'noteable']
> Extracted keywords: ['interleaf', 'yc']
message='OpenAI API response' path=https://api.openai.com/v1/embeddings processing_ms=25 request_id=95c73e9360e6473daab85cde93ca4c42 response_code=200

Node ID: 76d76348-52fb-49e6-95b8-2f7a3900fa1a
Similarity: None
Text: So this set me thinking. It was true that on my current trajectory, YC would be the last thing I …

Node ID: 61e1908a-79d2-426b-840e-926df469ac49
Similarity: None
Text: to the “YC GDP,” but as YC grows this becomes less and less of a joke. Now lots of startups get t…

Node ID: cac03004-5c02-4145-8e92-c320b1803847
Similarity: None
Text: stop there, of course, or you get merely photographic accuracy, and what makes a still life inter…

Node ID: f0d55e5e-5349-4243-ab01-d9dd7b12cd0a
Similarity: None
Text: of readers, but professional investors are thinking “Wow, that means they got all the returns.” B…

Node ID: 1516923c-0dee-4af2-b042-3e1f38de7e86
Similarity: None
Text: must tell readers things they don’t already know, and some people dislike being told such things…