docs for restructure

This commit is contained in:
Harrison Chase 2022-11-20 18:28:26 -08:00
parent 8c8eb47765
commit 71a0940435
28 changed files with 314 additions and 303 deletions

103
README.md
View File

@ -17,11 +17,6 @@ create a truly powerful app - the real power comes when you are able to
combine them with other sources of computation or knowledge.
This library is aimed at assisting in the development of those types of applications.
It aims to create:
1. a comprehensive collection of pieces you would ever want to combine
2. a flexible interface for combining pieces into a single comprehensive "chain"
3. a schema for easily saving and sharing those chains
## 📖 Documentation
@ -31,78 +26,42 @@ Please see [here](https://langchain.readthedocs.io/en/latest/?) for full documen
- Reference (full API docs)
- Resources (high level explanation of core concepts)
## 🚀 What can I do with this
## 🚀 What can this help with?
This project was largely inspired by a few projects seen on Twitter for which we thought it would make sense to have more explicit tooling. A lot of the initial functionality was done in an attempt to recreate those. Those are:
There are three main areas (with a forth coming soon) that LangChain is designed to help with.
These are, in increasing order of complexity:
1. LLM and Prompt usage
2. Chaining LLMs with other tools in a deterministic manner
3. Having a router LLM which uses other tools as needed
4. (Coming Soon) Memory
**[Self-ask-with-search](https://ofir.io/self-ask.pdf)**
### LLMs and Prompts
Calling out to an LLM once is pretty easy, with most of them being behind well documented APIs.
However, there are still some challenges going from that to an application running in production that LangChain attempts to address:
- Easy switching costs: by exposing a standard interface for all the top LLM providers, LangChain makes it easy to switch from one provider to another, whether it be for production use cases or just for testing stuff out.
- Prompt management: managing your prompts is easy when you only have one simple one, but can get tricky when you have a bunch or when they start to get more complex. LangChain provides a standard way for storing, constructing, and referencing prompts.
- Prompt optimization: despite the underlying models getting better and better, there is still currently a need for carefully constructing prompts.
- More coming soon
To recreate this paper, use the following code snippet or checkout the [example notebook](https://github.com/hwchase17/langchain/blob/master/docs/examples/demos/self_ask_with_search.ipynb).
### Chains
Using an LLM in isolation is fine for some simple applications, but many more complex ones require chaining LLMs - either with eachother or with other tools.
LangChain provides several parts to help with that:
- Standard interface for working with Chains
- Easy way to construct chains of LLMs
- Lots of integrations with other tools that you may want to use in conjunction with LLMs (search, databases, Python REPL, etc)
- End-to-end chains for common workflows (database question/answer, recursive summarization, etc)
```python
from langchain import SelfAskWithSearchChain, OpenAI, SerpAPIChain
### Routing Chains
Some applications will require not just a predetermined chain of calls to LLMs/other tools, but potentially an unknown chain that depends on the user input.
In these types of chains, there is a "router" LLM chain which has access to a suite of tools.
Depending on the user input, the router can then decide which, if any, of these tools to call.
To help develop applications like these, LangChain provides:
- Standard router and router chain interfaces
- Common router LLM chains from literature
- Common chains that can be used as tools
llm = OpenAI(temperature=0)
search = SerpAPIChain()
self_ask_with_search = SelfAskWithSearchChain(llm=llm, search_chain=search)
self_ask_with_search.run("What is the hometown of the reigning men's U.S. Open champion?")
```
**[LLM Math](https://twitter.com/amasad/status/1568824744367259648?s=20&t=-7wxpXBJinPgDuyHLouP1w)**
To recreate this example, use the following code snippet or check out the [example notebook](https://github.com/hwchase17/langchain/blob/master/docs/examples/demos/llm_math.ipynb).
```python
from langchain import OpenAI, LLMMathChain
llm = OpenAI(temperature=0)
llm_math = LLMMathChain(llm=llm)
llm_math.run("How many of the integers between 0 and 99 inclusive are divisible by 8?")
```
**Generic Prompting**
You can also use this for simple prompting pipelines, as in the below example and this [example notebook](https://github.com/hwchase17/langchain/blob/master/docs/examples/demos/simple_prompts.ipynb).
```python
from langchain import PromptTemplate, OpenAI, LLMChain
template = """Question: {question}
Answer: Let's think step by step."""
prompt = PromptTemplate(template=template, input_variables=["question"])
llm = OpenAI(temperature=0)
llm_chain = LLMChain(prompt=prompt, llm=llm)
question = "What NFL team won the Super Bowl in the year Justin Bieber was born?"
llm_chain.predict(question=question)
```
**Embed & Search Documents**
We support two vector databases to store and search embeddings -- FAISS and Elasticsearch. Here's a code snippet showing how to use FAISS to store embeddings and search for text similar to a query. Both database backends are featured in this [example notebook](https://github.com/hwchase17/langchain/blob/master/docs/examples/integrations/embeddings.ipynb).
```python
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.faiss import FAISS
from langchain.text_splitter import CharacterTextSplitter
with open('state_of_the_union.txt') as f:
state_of_the_union = f.read()
text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
texts = text_splitter.split_text(state_of_the_union)
embeddings = OpenAIEmbeddings()
docsearch = FAISS.from_texts(texts, embeddings)
query = "What did the president say about Ketanji Brown Jackson"
docs = docsearch.similarity_search(query)
```
### Memory
Coming soon.
## 🤖 Developer Guide

View File

@ -2,9 +2,24 @@ Demos
=====
The examples here are all end-to-end chains of specific applications.
They are separated into normal chains and then routing chains.
.. toctree::
:maxdepth: 1
:glob:
:caption: Chains
demos/*
demos/llm_math.ipynb
demos/map_reduce.ipynb
demos/simple_prompts.ipynb
demos/sqlite.ipynb
demos/vector_db_qa.ipynb
.. toctree::
:maxdepth: 1
:glob:
:caption: Routing Chains
demos/mrkl.ipynb
demos/react.ipynb
demos/self_ask_with_search.ipynb

View File

@ -40,10 +40,10 @@
"Thought 1:\u001b[32;1m\u001b[1;3m I need to search David Chanoff and find the U.S. Navy admiral he collaborated\n",
"with.\n",
"Action 1: Search[David Chanoff]\u001b[0m\n",
"Observation 1: \u001b[33;1m\u001b[1;3mDavid Chanoff is a noted author of non-fiction work. His work has typically involved collaborations with the principal protagonist of the work concerned. His collaborators have included; Augustus A. White, Joycelyn Elders, Đoàn Văn Toại, William J. Crowe, Ariel Sharon, Kenneth Good and Felix Zandman. He has also written about a wide range of subjects including literary history, education and foreign for The Washington Post, The New Republic and The New York Times Magazine. He has published more than twelve books.\u001b[0m\n",
"Observation 1: \u001b[36;1m\u001b[1;3mDavid Chanoff is a noted author of non-fiction work. His work has typically involved collaborations with the principal protagonist of the work concerned. His collaborators have included; Augustus A. White, Joycelyn Elders, Đoàn Văn Toại, William J. Crowe, Ariel Sharon, Kenneth Good and Felix Zandman. He has also written about a wide range of subjects including literary history, education and foreign for The Washington Post, The New Republic and The New York Times Magazine. He has published more than twelve books.\u001b[0m\n",
"Thought 2:\u001b[32;1m\u001b[1;3m The U.S. Navy admiral David Chanoff collaborated with is William J. Crowe.\n",
"Action 2: Search[William J. Crowe]\u001b[0m\n",
"Observation 2: \u001b[33;1m\u001b[1;3mWilliam James Crowe Jr. (January 2, 1925 October 18, 2007) was a United States Navy admiral and diplomat who served as the 11th chairman of the Joint Chiefs of Staff under Presidents Ronald Reagan and George H. W. Bush, and as the ambassador to the United Kingdom and Chair of the Intelligence Oversight Board under President Bill Clinton.\u001b[0m\n",
"Observation 2: \u001b[36;1m\u001b[1;3mWilliam James Crowe Jr. (January 2, 1925 October 18, 2007) was a United States Navy admiral and diplomat who served as the 11th chairman of the Joint Chiefs of Staff under Presidents Ronald Reagan and George H. W. Bush, and as the ambassador to the United Kingdom and Chair of the Intelligence Oversight Board under President Bill Clinton.\u001b[0m\n",
"Thought 3:\u001b[32;1m\u001b[1;3m William J. Crowe served as the ambassador to the United Kingdom under President Bill Clinton.\n",
"Action 3: Finish[Bill Clinton]\u001b[0m\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
@ -68,7 +68,7 @@
{
"cell_type": "code",
"execution_count": null,
"id": "2d6f4b1d",
"id": "3cb9d77c",
"metadata": {},
"outputs": [],
"source": []

View File

@ -8,12 +8,47 @@ create a truly powerful app - the real power comes when you are able to
combine them with other sources of computation or knowledge.
This library is aimed at assisting in the development of those types of applications.
It aims to create:
1. a comprehensive collection of pieces you would ever want to combine
2. a flexible interface for combining pieces into a single comprehensive "chain"
3. a schema for easily saving and sharing those chains
There are three main areas (with a forth coming soon) that LangChain is designed to help with.
These are, in increasing order of complexity:
1. LLM and Prompt usage
2. Chaining LLMs with other tools in a deterministic manner
3. Having a router LLM which uses other tools as needed
4. (Coming Soon) Memory
**LLMs and Prompts**
Calling out to an LLM once is pretty easy, with most of them being behind well documented APIs.
However, there are still some challenges going from that to an application running in production that LangChain attempts to address:
- Easy switching costs: by exposing a standard interface for all the top LLM providers, LangChain makes it easy to switch from one provider to another, whether it be for production use cases or just for testing stuff out.
- Prompt management: managing your prompts is easy when you only have one simple one, but can get tricky when you have a bunch or when they start to get more complex. LangChain provides a standard way for storing, constructing, and referencing prompts.
- Prompt optimization: despite the underlying models getting better and better, there is still currently a need for carefully constructing prompts.
- More coming soon
**Chains**
Using an LLM in isolation is fine for some simple applications, but many more complex ones require chaining LLMs - either with eachother or with other tools.
LangChain provides several parts to help with that:
- Standard interface for working with Chains
- Easy way to construct chains of LLMs
- Lots of integrations with other tools that you may want to use in conjunction with LLMs (search, databases, Python REPL, etc)
- End-to-end chains for common workflows (database question/answer, recursive summarization, etc)
**Routing Chains**
Some applications will require not just a predetermined chain of calls to LLMs/other tools, but potentially an unknown chain that depends on the user input.
In these types of chains, there is a "router" LLM chain which has access to a suite of tools.
Depending on the user input, the router can then decide which, if any, of these tools to call.
To help develop applications like these, LangChain provides:
- Standard router and router chain interfaces
- Common router LLM chains from literature
- Common chains that can be used as tools
**Memory**
Coming soon.
Documentation Structure
=======================
The documentation is structured into the following sections:
@ -62,6 +97,7 @@ common tasks or cool demos.
modules/text_splitter
modules/vectorstore
modules/chains
modules/routing_chains
Full API documentation. This is the place to look if you want to

View File

@ -0,0 +1,7 @@
:mod:`langchain.routing_chains`
===============================
.. automodule:: langchain.routing_chains
:members:
:undoc-members:

View File

@ -21,7 +21,7 @@ from langchain.prompts import (
Prompt,
PromptTemplate,
)
from langchain.smart_chains import MRKLChain, ReActChain, SelfAskWithSearchChain
from langchain.routing_chains import MRKLChain, ReActChain, SelfAskWithSearchChain
from langchain.sql_database import SQLDatabase
from langchain.vectorstores import FAISS, ElasticVectorSearch

View File

@ -0,0 +1,14 @@
"""Routing chains."""
from langchain.routing_chains.mrkl.base import MRKLChain
from langchain.routing_chains.react.base import ReActChain
from langchain.routing_chains.router import LLMRouter
from langchain.routing_chains.routing_chain import RoutingChain
from langchain.routing_chains.self_ask_with_search.base import SelfAskWithSearchChain
__all__ = [
"MRKLChain",
"SelfAskWithSearchChain",
"ReActChain",
"LLMRouter",
"RoutingChain",
]

View File

@ -4,9 +4,9 @@ from typing import Any, Callable, List, NamedTuple, Optional, Tuple
from langchain.chains.llm import LLMChain
from langchain.llms.base import LLM
from langchain.prompts import PromptTemplate
from langchain.smart_chains.mrkl.prompt import BASE_TEMPLATE
from langchain.smart_chains.router import LLMRouterChain
from langchain.smart_chains.router_expert import ExpertConfig, RouterExpertChain
from langchain.routing_chains.mrkl.prompt import BASE_TEMPLATE
from langchain.routing_chains.router import LLMRouter
from langchain.routing_chains.routing_chain import RoutingChain, ToolConfig
FINAL_ANSWER_ACTION = "Final Answer: "
@ -46,7 +46,7 @@ def get_action_and_input(llm_output: str) -> Tuple[str, str]:
return action, action_input.strip(" ").strip('"')
class MRKLRouterChain(LLMRouterChain):
class MRKLRouterChain(LLMRouter):
"""Router for the MRKL chain."""
@property
@ -71,11 +71,11 @@ class MRKLRouterChain(LLMRouterChain):
stops = ["\nObservation"]
super().__init__(llm_chain=llm_chain, stops=stops, **kwargs)
def _extract_action_and_input(self, text: str) -> Optional[Tuple[str, str]]:
def _extract_tool_and_input(self, text: str) -> Optional[Tuple[str, str]]:
return get_action_and_input(text)
class MRKLChain(RouterExpertChain):
class MRKLChain(RoutingChain):
"""Chain that implements the MRKL system.
Example:
@ -130,6 +130,6 @@ class MRKLChain(RouterExpertChain):
"""
router_chain = MRKLRouterChain(llm, chains)
expert_configs = [
ExpertConfig(expert_name=c.action_name, expert=c.action) for c in chains
ToolConfig(tool_name=c.action_name, tool=c.action) for c in chains
]
return cls(router_chain=router_chain, expert_configs=expert_configs, **kwargs)

View File

@ -8,12 +8,12 @@ from langchain.chains.llm import LLMChain
from langchain.docstore.base import Docstore
from langchain.docstore.document import Document
from langchain.llms.base import LLM
from langchain.smart_chains.react.prompt import PROMPT
from langchain.smart_chains.router import LLMRouterChain
from langchain.smart_chains.router_expert import ExpertConfig, RouterExpertChain
from langchain.routing_chains.react.prompt import PROMPT
from langchain.routing_chains.router import LLMRouter
from langchain.routing_chains.routing_chain import RoutingChain, ToolConfig
class ReActRouterChain(LLMRouterChain, BaseModel):
class ReActRouterChain(LLMRouter, BaseModel):
"""Router for the ReAct chin."""
i: int = 1
@ -27,7 +27,7 @@ class ReActRouterChain(LLMRouterChain, BaseModel):
def _fix_text(self, text: str) -> str:
return text + f"\nAction {self.i}:"
def _extract_action_and_input(self, text: str) -> Optional[Tuple[str, str]]:
def _extract_tool_and_input(self, text: str) -> Optional[Tuple[str, str]]:
action_prefix = f"Action {self.i}: "
if not text.split("\n")[-1].startswith(action_prefix):
return None
@ -83,7 +83,7 @@ class DocstoreExplorer:
return self.document.lookup(term)
class ReActChain(RouterExpertChain):
class ReActChain(RoutingChain):
"""Chain that implements the ReAct paper.
Example:
@ -95,12 +95,10 @@ class ReActChain(RouterExpertChain):
def __init__(self, llm: LLM, docstore: Docstore, **kwargs: Any):
"""Initialize with the LLM and a docstore."""
router_chain = ReActRouterChain(llm)
router = ReActRouterChain(llm)
docstore_explorer = DocstoreExplorer(docstore)
expert_configs = [
ExpertConfig(expert_name="Search", expert=docstore_explorer.search),
ExpertConfig(expert_name="Lookup", expert=docstore_explorer.lookup),
tool_configs = [
ToolConfig(tool_name="Search", tool=docstore_explorer.search),
ToolConfig(tool_name="Lookup", tool=docstore_explorer.lookup),
]
super().__init__(
router_chain=router_chain, expert_configs=expert_configs, **kwargs
)
super().__init__(router=router, expert_configs=tool_configs, **kwargs)

View File

@ -0,0 +1,86 @@
"""Chain that takes in an input and produces an action and action input."""
from abc import ABC, abstractmethod
from typing import NamedTuple, Optional, Tuple
from pydantic import BaseModel
from langchain.chains.llm import LLMChain
class RouterOutput(NamedTuple):
"""Output of a router."""
tool: str
tool_input: str
log: str
class Router(ABC):
"""Chain responsible for deciding the action to take."""
@abstractmethod
def route(self, text: str) -> RouterOutput:
"""Given input, decided how to route it.
Args:
text: input string
Returns:
RouterOutput specifying what tool to use.
"""
@property
@abstractmethod
def observation_prefix(self) -> str:
"""Prefix to append the observation with."""
@property
@abstractmethod
def router_prefix(self) -> str:
"""Prefix to append the router call with."""
@property
def finish_tool_name(self) -> str:
"""Name of the tool to use to finish the chain."""
return "Final Answer"
@property
def starter_string(self) -> str:
"""Put this string after user input but before first router call."""
return "\n"
class LLMRouter(Router, BaseModel, ABC):
"""Router that uses an LLM."""
llm_chain: LLMChain
@abstractmethod
def _extract_tool_and_input(self, text: str) -> Optional[Tuple[str, str]]:
"""Extract tool and tool input from llm output."""
def _fix_text(self, text: str) -> str:
"""Fix the text."""
raise ValueError("fix_text not implemented for this router.")
def route(self, text: str) -> RouterOutput:
"""Given input, decided how to route it.
Args:
text: input string
Returns:
RouterOutput specifying what tool to use.
"""
input_key = self.llm_chain.input_keys[0]
inputs = {input_key: text, "stop": [self.observation_prefix]}
full_output = self.llm_chain.predict(**inputs)
parsed_output = self._extract_tool_and_input(full_output)
while parsed_output is None:
full_output = self._fix_text(full_output)
inputs = {input_key: text + full_output, "stop": [self.observation_prefix]}
output = self.llm_chain.predict(**inputs)
full_output += output
parsed_output = self._extract_tool_and_input(full_output)
tool, tool_input = parsed_output
return RouterOutput(tool, tool_input, full_output)

View File

@ -0,0 +1,70 @@
"""Router-Expert framework."""
from typing import Callable, Dict, List, NamedTuple
from pydantic import BaseModel, Extra
from langchain.chains.base import Chain
from langchain.input import ChainedInput, get_color_mapping
from langchain.routing_chains.router import Router
class ToolConfig(NamedTuple):
"""Configuration for tools."""
tool_name: str
tool: Callable[[str], str]
class RoutingChain(Chain, BaseModel):
"""Chain that uses a router to use tools."""
router: Router
"""Router to use."""
tool_configs: List[ToolConfig]
"""Tool configs this chain has access to."""
input_key: str = "question" #: :meta private:
output_key: str = "answer" #: :meta private:
class Config:
"""Configuration for this pydantic object."""
extra = Extra.forbid
arbitrary_types_allowed = True
@property
def input_keys(self) -> List[str]:
"""Expect input key.
:meta private:
"""
return [self.input_key]
@property
def output_keys(self) -> List[str]:
"""Expect output key.
:meta private:
"""
return [self.output_key]
def _call(self, inputs: Dict[str, str]) -> Dict[str, str]:
name_to_tool_map = {tc.tool_name: tc.tool for tc in self.tool_configs}
starter_string = (
inputs[self.input_key]
+ self.router.starter_string
+ self.router.router_prefix
)
chained_input = ChainedInput(starter_string, verbose=self.verbose)
color_mapping = get_color_mapping(
[c.tool_name for c in self.tool_configs], excluded_colors=["green"]
)
while True:
output = self.router.route(chained_input.input)
chained_input.add(output.log, color="green")
if output.tool == self.router.finish_tool_name:
return {self.output_key: output.tool_input}
chain = name_to_tool_map[output.tool]
observation = chain(output.tool_input)
chained_input.add(f"\n{self.router.observation_prefix}")
chained_input.add(observation, color=color_mapping[output.tool])
chained_input.add(f"\n{self.router.router_prefix}")

View File

@ -4,12 +4,12 @@ from typing import Any, Tuple
from langchain.chains.llm import LLMChain
from langchain.chains.serpapi import SerpAPIChain
from langchain.llms.base import LLM
from langchain.smart_chains.router import LLMRouterChain
from langchain.smart_chains.router_expert import ExpertConfig, RouterExpertChain
from langchain.smart_chains.self_ask_with_search.prompt import PROMPT
from langchain.routing_chains.router import LLMRouter
from langchain.routing_chains.routing_chain import RoutingChain, ToolConfig
from langchain.routing_chains.self_ask_with_search.prompt import PROMPT
class SelfAskWithSearchRouter(LLMRouterChain):
class SelfAskWithSearchRouter(LLMRouter):
"""Router for the self-ask-with-search paper."""
def __init__(self, llm: LLM, **kwargs: Any):
@ -17,7 +17,7 @@ class SelfAskWithSearchRouter(LLMRouterChain):
llm_chain = LLMChain(llm=llm, prompt=PROMPT)
super().__init__(llm_chain=llm_chain, **kwargs)
def _extract_action_and_input(self, text: str) -> Tuple[str, str]:
def _extract_tool_and_input(self, text: str) -> Tuple[str, str]:
followup = "Follow up:"
if "\n" not in text:
last_line = text
@ -52,8 +52,13 @@ class SelfAskWithSearchRouter(LLMRouterChain):
"""Prefix to append the router call with."""
return ""
@property
def starter_string(self) -> str:
"""Put this string after user input but before first router call."""
return "\nAre follow up questions needed here:"
class SelfAskWithSearchChain(RouterExpertChain):
class SelfAskWithSearchChain(RoutingChain):
"""Chain that does self ask with search.
Example:
@ -68,12 +73,6 @@ class SelfAskWithSearchChain(RouterExpertChain):
"""Initialize with just an LLM and a search chain."""
intermediate = "\nIntermediate answer:"
router = SelfAskWithSearchRouter(llm, stops=[intermediate])
expert_configs = [
ExpertConfig(expert_name="Intermediate Answer", expert=search_chain.run)
]
super().__init__(
router_chain=router,
expert_configs=expert_configs,
starter_string="\nAre follow up questions needed here:",
**kwargs
)
search_tool = ToolConfig(tool_name="Intermediate Answer", tool=search_chain.run)
expert_configs = [search_tool]
super().__init__(router=router, expert_configs=expert_configs, **kwargs)

View File

@ -1,6 +0,0 @@
"""Smart chains."""
from langchain.smart_chains.mrkl.base import MRKLChain
from langchain.smart_chains.react.base import ReActChain
from langchain.smart_chains.self_ask_with_search.base import SelfAskWithSearchChain
__all__ = ["MRKLChain", "SelfAskWithSearchChain", "ReActChain"]

View File

@ -1,90 +0,0 @@
"""Chain that takes in an input and produces an action and action input."""
from abc import ABC, abstractmethod
from typing import Dict, List, Optional, Tuple
from pydantic import BaseModel
from langchain.chains.base import Chain
from langchain.chains.llm import LLMChain
class RouterChain(Chain, BaseModel, ABC):
"""Chain responsible for deciding the action to take."""
input_key: str = "input_text" #: :meta private:
action_key: str = "action" #: :meta private:
action_input_key: str = "action_input" #: :meta private:
log_key: str = "log" #: :meta private:
@property
def input_keys(self) -> List[str]:
"""Will be the input key.
:meta private:
"""
return [self.input_key]
@property
def output_keys(self) -> List[str]:
"""Return three keys: the action, the action input, and the log.
:meta private:
"""
return [self.action_key, self.action_input_key, self.log_key]
@abstractmethod
def get_action_and_input(self, text: str) -> Tuple[str, str, str]:
"""Return action, action input, and log (in that order)."""
@property
@abstractmethod
def observation_prefix(self) -> str:
"""Prefix to append the observation with."""
@property
@abstractmethod
def router_prefix(self) -> str:
"""Prefix to append the router call with."""
@property
def finish_action_name(self) -> str:
"""Name of the action of when to finish the chain."""
return "Final Answer"
def _call(self, inputs: Dict[str, str]) -> Dict[str, str]:
action, action_input, log = self.get_action_and_input(inputs[self.input_key])
return {
self.action_key: action,
self.action_input_key: action_input,
self.log_key: log,
}
class LLMRouterChain(RouterChain, BaseModel, ABC):
"""RouterChain that uses an LLM."""
llm_chain: LLMChain
stops: Optional[List[str]]
@abstractmethod
def _extract_action_and_input(self, text: str) -> Optional[Tuple[str, str]]:
"""Extract action and action input from llm output."""
def _fix_text(self, text: str) -> str:
"""Fix the text."""
raise ValueError("fix_text not implemented for this router.")
def get_action_and_input(self, text: str) -> Tuple[str, str, str]:
"""Return action, action input, and log (in that order)."""
input_key = self.llm_chain.input_keys[0]
inputs = {input_key: text, "stop": self.stops}
full_output = self.llm_chain.predict(**inputs)
parsed_output = self._extract_action_and_input(full_output)
while parsed_output is None:
full_output = self._fix_text(full_output)
inputs = {input_key: text + full_output, "stop": self.stops}
output = self.llm_chain.predict(**inputs)
full_output += output
parsed_output = self._extract_action_and_input(full_output)
action, action_input = parsed_output
return action, action_input, full_output

View File

@ -1,77 +0,0 @@
"""Router-Expert framework."""
from typing import Callable, Dict, List, NamedTuple
from pydantic import BaseModel, Extra
from langchain.chains.base import Chain
from langchain.input import ChainedInput, get_color_mapping
from langchain.smart_chains.router import RouterChain
class ExpertConfig(NamedTuple):
"""Configuration for experts."""
expert_name: str
expert: Callable[[str], str]
class RouterExpertChain(Chain, BaseModel):
"""Chain that implements the Router/Expert system."""
router_chain: RouterChain
"""Router chain."""
expert_configs: List[ExpertConfig]
"""Expert configs this chain has access to."""
starter_string: str = "\n"
"""String to put after user input but before first router."""
input_key: str = "question" #: :meta private:
output_key: str = "answer" #: :meta private:
class Config:
"""Configuration for this pydantic object."""
extra = Extra.forbid
arbitrary_types_allowed = True
@property
def input_keys(self) -> List[str]:
"""Expect input key.
:meta private:
"""
return [self.input_key]
@property
def output_keys(self) -> List[str]:
"""Expect output key.
:meta private:
"""
return [self.output_key]
def _call(self, inputs: Dict[str, str]) -> Dict[str, str]:
action_to_chain_map = {e.expert_name: e.expert for e in self.expert_configs}
starter_string = (
inputs[self.input_key]
+ self.starter_string
+ self.router_chain.router_prefix
)
chained_input = ChainedInput(
starter_string,
verbose=self.verbose,
)
color_mapping = get_color_mapping(
[c.expert_name for c in self.expert_configs], excluded_colors=["green"]
)
while True:
action, action_input, log = self.router_chain.get_action_and_input(
chained_input.input
)
chained_input.add(log, color="green")
if action == self.router_chain.finish_action_name:
return {self.output_key: action_input}
chain = action_to_chain_map[action]
ca = chain(action_input)
chained_input.add(f"\n{self.router_chain.observation_prefix}")
chained_input.add(ca, color=color_mapping[action])
chained_input.add(f"\n{self.router_chain.router_prefix}")

View File

@ -2,7 +2,7 @@
from langchain.docstore.wikipedia import Wikipedia
from langchain.llms.openai import OpenAI
from langchain.smart_chains.react.base import ReActChain
from langchain.routing_chains.react.base import ReActChain
def test_react() -> None:

View File

@ -1,7 +1,7 @@
"""Integration test for self ask with search."""
from langchain.chains.serpapi import SerpAPIChain
from langchain.llms.openai import OpenAI
from langchain.smart_chains.self_ask_with_search.base import SelfAskWithSearchChain
from langchain.routing_chains.self_ask_with_search.base import SelfAskWithSearchChain
def test_self_ask_with_search() -> None:

View File

@ -0,0 +1 @@
"""Test routing chain functionality."""

View File

@ -3,12 +3,12 @@
import pytest
from langchain.prompts import PromptTemplate
from langchain.smart_chains.mrkl.base import (
from langchain.routing_chains.mrkl.base import (
ChainConfig,
MRKLRouterChain,
get_action_and_input,
)
from langchain.smart_chains.mrkl.prompt import BASE_TEMPLATE
from langchain.routing_chains.mrkl.prompt import BASE_TEMPLATE
from tests.unit_tests.llms.fake_llm import FakeLLM

View File

@ -8,7 +8,7 @@ from langchain.docstore.base import Docstore
from langchain.docstore.document import Document
from langchain.llms.base import LLM
from langchain.prompts.prompt import PromptTemplate
from langchain.smart_chains.react.base import ReActChain, ReActRouterChain
from langchain.routing_chains.react.base import ReActChain, ReActRouterChain
_PAGE_CONTENT = """This is a page about LangChain.
@ -53,10 +53,10 @@ def test_predict_until_observation_normal() -> None:
outputs = ["foo\nAction 1: search[foo]"]
fake_llm = FakeListLLM(outputs)
router_chain = ReActRouterChain(llm=fake_llm)
action, directive, ret_text = router_chain.get_action_and_input("")
assert ret_text == outputs[0]
assert action == "search"
assert directive == "foo"
output = router_chain.route("")
assert output.log == outputs[0]
assert output.tool == "search"
assert output.tool_input == "foo"
def test_predict_until_observation_repeat() -> None:
@ -64,10 +64,10 @@ def test_predict_until_observation_repeat() -> None:
outputs = ["foo", " search[foo]"]
fake_llm = FakeListLLM(outputs)
router_chain = ReActRouterChain(llm=fake_llm)
action, directive, ret_text = router_chain.get_action_and_input("")
assert ret_text == "foo\nAction 1: search[foo]"
assert action == "search"
assert directive == "foo"
output = router_chain.route("")
assert output.log == "foo\nAction 1: search[foo]"
assert output.tool == "search"
assert output.tool_input == "foo"
def test_react_chain() -> None:

View File

@ -1 +0,0 @@
"""Test smart chain functionality."""