Streaming With LangChain
Streaming is critical in making applications based on LLMs feel responsive to end-users.
Important LangChain primitives like LLMs, parsers, prompts, retrievers, and agents implement the LangChain Runnable Interface.
This interface provides two general approaches to stream content:
- sync
stream
and asyncastream
: a default implementation of streaming that streams the final output from the chain. - async
astream_events
and asyncastream_log
: these provide a way to stream both intermediate steps and final output from the chain.
Let's take a look at both approaches, and try to understand a how to use them. 🥷
Using Stream
All Runnable
objects implement a sync method called stream
and an async variant called astream
.
These methods are designed to stream the final output in chunks, yielding each chunk as soon as it is available.
Streaming is only possible if all steps in the program know how to process an input stream; i.e., process an input chunk one at a time, and yield a corresponding output chunk.
The complexity of this processing can vary, from straightforward tasks like emitting tokens produced by an LLM, to more challenging ones like streaming parts of JSON results before the entire JSON is complete.
The best place to start exploring streaming is with the single most important components in LLMs apps-- the LLMs themselves!
LLMs and Chat Models
Large language models and their chat variants are the primary bottleneck in LLM based apps. 🙊
Large language models can take several seconds to generate a complete response to a query. This is far slower than the ~200-300 ms threshold at which an application feels responsive to an end user.
The key strategy to make the application feel more responsive is to show intermediate progress; e.g., to stream the output from the model token by token.
# Showing the example using anthropic, but you can use
# your favorite chat model!
from langchain.chat_models import ChatAnthropic
model = ChatAnthropic()
chunks = []
async for chunk in model.astream("hello. tell me something about yourself"):
chunks.append(chunk)
print(chunk.content, end="|", flush=True)
Hello|!| My| name| is| Claude|.| I|'m| an| AI| assistant| created| by| An|throp|ic| to| be| helpful|,| harmless|,| and| honest|.||
Let's inspect one of the chunks
chunks[0]
AIMessageChunk(content=' Hello')
We got back something called an AIMessageChunk
. This chunk represents a part of an AIMessage
.
Message chunks are additive by design -- one can simply add them up to get the state of the response so far!
chunks[0] + chunks[1] + chunks[2] + chunks[3] + chunks[4]
AIMessageChunk(content=' Hello! My name is')
Chains
Virtually all LLM applications involve more steps than just a call to a language model.
Let's build a simple chain using LangChain Expression Language
(LCEL
) that combines a prompt, model and a parser and verify that streaming works.
We will use StrOutputParser
to parse the output from the model. This is a simple parser that extracts the content
field from an AIMessageChunk
, giving us the token
returned by the model.
:::{.callout-tip}
LCEL is a declarative way to specify a "program" by chainining together different LangChain primitives. Chains created using LCEL benefit from an automatic implementation of stream
, and astream
allowing streaming of the final output. In fact, chains created with LCEL implement the entire standard Runnable interface.
:::
from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
prompt = ChatPromptTemplate.from_template("tell me a joke about {topic}")
parser = StrOutputParser()
chain = prompt | model | parser
async for chunk in chain.astream({"topic": "parrot"}):
print(chunk, end="|", flush=True)
Here|'s| a| silly| joke| about| a| par|rot|:|
What| kind| of| teacher| gives| good| advice|?| An| ap|-|parent| (|app|arent|)| one|!||
:::{.callout-note}
You do not have to use the LangChain Expression Language
to use LangChain and can instead rely on a standard imperative programming approach by
caling invoke
, batch
or stream
on each component individually, assigning the results to variables and then using them downstream as you see fit.
If that works for your needs, then that's fine by us 👌! :::
Working with Input Streams
What if you wanted to stream JSON from the output as it was being generated?
If you were to rely on json.loads
to parse the partial json, the parsing would fail as the partial json wouldn't be valid json.
You'd likely be at a complete loss of what to do and claim that it wasn't possible to stream JSON.
Well, turns out there is a way to do it -- the parser needs to operate on the input stream, and attempt to "auto-complete" the partial json into a valid state.
Let's see such a parser in action to understand what this means.
from langchain_core.output_parsers import JsonOutputParser
chain = (
model | JsonOutputParser()
) # Due to a bug in older versions of Langchain, JsonOutputParser did not stream results from some models
async for text in chain.astream(
'output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`'
):
print(text, flush=True)
{}
{'countries': []}
{'countries': [{}]}
{'countries': [{'name': ''}]}
{'countries': [{'name': 'France'}]}
{'countries': [{'name': 'France', 'population': 67}]}
{'countries': [{'name': 'France', 'population': 6739}]}
{'countries': [{'name': 'France', 'population': 673915}]}
{'countries': [{'name': 'France', 'population': 67391582}]}
{'countries': [{'name': 'France', 'population': 67391582}, {}]}
{'countries': [{'name': 'France', 'population': 67391582}, {'name': ''}]}
{'countries': [{'name': 'France', 'population': 67391582}, {'name': 'Sp'}]}
{'countries': [{'name': 'France', 'population': 67391582}, {'name': 'Spain'}]}
{'countries': [{'name': 'France', 'population': 67391582}, {'name': 'Spain', 'population': 46}]}
{'countries': [{'name': 'France', 'population': 67391582}, {'name': 'Spain', 'population': 4675}]}
{'countries': [{'name': 'France', 'population': 67391582}, {'name': 'Spain', 'population': 467547}]}
{'countries': [{'name': 'France', 'population': 67391582}, {'name': 'Spain', 'population': 46754778}]}
{'countries': [{'name': 'France', 'population': 67391582}, {'name': 'Spain', 'population': 46754778}, {}]}
{'countries': [{'name': 'France', 'population': 67391582}, {'name': 'Spain', 'population': 46754778}, {'name': ''}]}
{'countries': [{'name': 'France', 'population': 67391582}, {'name': 'Spain', 'population': 46754778}, {'name': 'Japan'}]}
{'countries': [{'name': 'France', 'population': 67391582}, {'name': 'Spain', 'population': 46754778}, {'name': 'Japan', 'population': 12}]}
{'countries': [{'name': 'France', 'population': 67391582}, {'name': 'Spain', 'population': 46754778}, {'name': 'Japan', 'population': 12647}]}
{'countries': [{'name': 'France', 'population': 67391582}, {'name': 'Spain', 'population': 46754778}, {'name': 'Japan', 'population': 1264764}]}
{'countries': [{'name': 'France', 'population': 67391582}, {'name': 'Spain', 'population': 46754778}, {'name': 'Japan', 'population': 126476461}]}
Now, let's break streaming. We'll use the previous example and append an extraction function at the end that extracts the country names from the finalized JSON.
:::{.callout-warning}
Any steps in the chain that operate on finalized inputs rather than on input streams can break streaming functionality via stream
or astream
.
:::
:::{.callout-tip}
Later, we will discuss the astream_events
API which streams results from intermediate steps. This API will stream results from intermediate steps even if the chain contains steps that only operate on finalized inputs.
:::
from langchain_core.output_parsers import (
JsonOutputParser,
)
# A function that operates on finalized inputs
# rather than on an input_stream
def _extract_country_names(inputs):
"""A function that does not operates on input streams and breaks streaming."""
if not isinstance(inputs, dict):
return ""
if "countries" not in inputs:
return ""
countries = inputs["countries"]
if not isinstance(countries, list):
return ""
country_names = [
country.get("name") for country in countries if isinstance(country, dict)
]
return country_names
chain = model | JsonOutputParser() | _extract_country_names
async for text in chain.astream(
'output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`'
):
print(text, end="|", flush=True)
['France', 'Spain', 'Japan']|
Generator Functions
Le'ts fix the streaming using a generator function that can operate on the input stream.
:::{.callout-tip}
A generator function (a function that uses yield
) allows writing code that operators on input streams
:::
from langchain_core.output_parsers import JsonOutputParser
async def _extract_country_names_streaming(input_stream):
"""A function that operates on input streams."""
country_names_so_far = set()
async for input in input_stream:
if not isinstance(input, dict):
continue
if "countries" not in input:
continue
countries = input["countries"]
if not isinstance(countries, list):
continue
for country in countries:
name = country.get("name")
if not name:
continue
if name not in country_names_so_far:
yield name
country_names_so_far.add(name)
chain = model | JsonOutputParser() | _extract_country_names_streaming
async for text in chain.astream(
'output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`'
):
print(text, end="|", flush=True)
France|Sp|Spain|Japan|
:::{.callout-note}
Because the code above is relying on JSON auto-completion, you may see partial names of countries (e.g., Sp
and Spain
), which is not what one would want for an extraction result!
We're focusing on streaming concepts, not necessarily the results of the chains. :::
Non-streaming components
Some built-in components like Retrievers do not offer any streaming
. What happens if we try to stream
them? 🤨
from langchain_community.vectorstores import FAISS
from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables import RunnablePassthrough
from langchain_openai import OpenAIEmbeddings
template = """Answer the question based only on the following context:
{context}
Question: {question}
"""
prompt = ChatPromptTemplate.from_template(template)
vectorstore = FAISS.from_texts(
["harrison worked at kensho", "harrison likes spicy food"],
embedding=OpenAIEmbeddings(),
)
retriever = vectorstore.as_retriever()
chunks = [chunk for chunk in retriever.stream("where did harrison work?")]
chunks
[[Document(page_content='harrison worked at kensho'),
Document(page_content='harrison likes spicy food')]]
Stream just yielded the final result from that component.
This is OK 🥹! Not all components have to implement streaming -- in some cases streaming is either unnecessary, difficult or just doesn't make sense.
:::{.callout-tip} An LCEL chain constructed using using non-streaming components, will still be able to stream in a lot of cases, with streaming of partial output starting after the last non-streaming step in the chain. :::
retrieval_chain = (
{
"context": retriever.with_config(run_name="Docs"),
"question": RunnablePassthrough(),
}
| prompt
| model
| StrOutputParser()
)
for chunk in retrieval_chain.stream(
"Where did harrison work? " "Write 3 made up sentences about this place."
):
print(chunk, end="|", flush=True)
Based| on| the| given| context|,| the| only| information| provided| about| where| Harrison| worked| is| that| he| worked| at| Ken|sh|o|.| Since| there| are| no| other| details| provided| about| Ken|sh|o|,| I| do| not| have| enough| information| to| write| 3| additional| made| up| sentences| about| this| place|.| I| can| only| state| that| Harrison| worked| at| Ken|sh|o|.||
Now that we've seen how stream
and astream
work, let's venture into the world of streaming events. 🏞️
Using Stream Events
Event Streaming is a beta API. This API may change a bit based on feedback.
:::{.callout-note} Introduced in langchain-core 0.1.14. :::
import langchain_core
langchain_core.__version__
'0.1.18'
For the astream_events
API to work properly:
- Use
async
throughout the code to the extent possible (e.g., async tools etc) - Propagate callbacks if defining custom functions / runnables
- Whenever using runnables without LCEL, make sure to call
.astream()
on LLMs rather than.ainvoke
to force the LLM to stream tokens. - Let us know if anything doesn't work as expected! :)
Event Reference
Below is a reference table that shows some events that might be emitted by the various Runnable objects.
:::{.callout-note}
When streaming is implemented properly, the inputs to a runnable will not be known until after the input stream has been entirely consumed. This means that inputs
will often be included only for end
events and rather than for start
events.
:::
event | name | chunk | input | output |
---|---|---|---|---|
on_chat_model_start | [model name] | {"messages": [[SystemMessage, HumanMessage]]} | ||
on_chat_model_stream | [model name] | AIMessageChunk(content="hello") | ||
on_chat_model_end | [model name] | {"messages": [[SystemMessage, HumanMessage]]} | {"generations": [...], "llm_output": None, ...} | |
on_llm_start | [model name] | {'input': 'hello'} | ||
on_llm_stream | [model name] | 'Hello' | ||
on_llm_end | [model name] | 'Hello human!' | ||
on_chain_start | format_docs | |||
on_chain_stream | format_docs | "hello world!, goodbye world!" | ||
on_chain_end | format_docs | [Document(...)] | "hello world!, goodbye world!" | |
on_tool_start | some_tool | {"x": 1, "y": "2"} | ||
on_tool_stream | some_tool | {"x": 1, "y": "2"} | ||
on_tool_end | some_tool | {"x": 1, "y": "2"} | ||
on_retriever_start | [retriever name] | {"query": "hello"} | ||
on_retriever_chunk | [retriever name] | {documents: [...]} | ||
on_retriever_end | [retriever name] | {"query": "hello"} | {documents: [...]} | |
on_prompt_start | [template_name] | {"question": "hello"} | ||
on_prompt_end | [template_name] | {"question": "hello"} | ChatPromptValue(messages: [SystemMessage, ...]) |
Chat Model
Let's start off by looking at the events produced by a chat model.
events = []
async for event in model.astream_events("hello", version="v1"):
events.append(event)
/home/eugene/src/langchain/libs/core/langchain_core/_api/beta_decorator.py:86: LangChainBetaWarning: This API is in beta and may change in the future.
warn_beta(
:::{.callout-note}
Hey what's that funny version="v1" parameter in the API?! 😾
This is a beta API, and we're almost certainly going to make some changes to it.
This version parameter will allow us to mimimize such breaking changes to your code.
In short, we are annoying you now, so we don't have to annoy you later. ::: Let's take a look at the few of the start event and a few of the end events.
events[:3]
[{'event': 'on_chat_model_start',
'run_id': '555843ed-3d24-4774-af25-fbf030d5e8c4',
'name': 'ChatAnthropic',
'tags': [],
'metadata': {},
'data': {'input': 'hello'}},
{'event': 'on_chat_model_stream',
'run_id': '555843ed-3d24-4774-af25-fbf030d5e8c4',
'tags': [],
'metadata': {},
'name': 'ChatAnthropic',
'data': {'chunk': AIMessageChunk(content=' Hello')}},
{'event': 'on_chat_model_stream',
'run_id': '555843ed-3d24-4774-af25-fbf030d5e8c4',
'tags': [],
'metadata': {},
'name': 'ChatAnthropic',
'data': {'chunk': AIMessageChunk(content='!')}}]
events[-2:]
[{'event': 'on_chat_model_stream',
'run_id': '555843ed-3d24-4774-af25-fbf030d5e8c4',
'tags': [],
'metadata': {},
'name': 'ChatAnthropic',
'data': {'chunk': AIMessageChunk(content='')}},
{'event': 'on_chat_model_end',
'name': 'ChatAnthropic',
'run_id': '555843ed-3d24-4774-af25-fbf030d5e8c4',
'tags': [],
'metadata': {},
'data': {'output': AIMessageChunk(content=' Hello!')}}]
Chain
Let's revisit the example chain that parsed streaming JSON to explore the streaming events API.
chain = (
model | JsonOutputParser()
) # Due to a bug in older versions of Langchain, JsonOutputParser did not stream results from some models
events = [
event
async for event in chain.astream_events(
'output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`',
version="v1",
)
]
If you examine at the first few events, you'll notice that there are 3 different start events rather than 2 start events.
The three start events correspond to:
- The chain (model + parser)
- The model
- The parser
events[:3]
[{'event': 'on_chain_start',
'run_id': 'b1074bff-2a17-458b-9e7b-625211710df4',
'name': 'RunnableSequence',
'tags': [],
'metadata': {},
'data': {'input': 'output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`'}},
{'event': 'on_chat_model_start',
'name': 'ChatAnthropic',
'run_id': '6072be59-1f43-4f1c-9470-3b92e8406a99',
'tags': ['seq:step:1'],
'metadata': {},
'data': {'input': {'messages': [[HumanMessage(content='output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`')]]}}},
{'event': 'on_parser_start',
'name': 'JsonOutputParser',
'run_id': 'bf978194-0eda-4494-ad15-3a5bfe69cd59',
'tags': ['seq:step:2'],
'metadata': {},
'data': {}}]
What do you think you'd see if you looked at the last 3 events? what about the middle? Let's use this API to take output the stream events from the model and the parser. We're ignoring start events, end events and events from the chain.
num_events = 0
async for event in chain.astream_events(
'output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`',
version="v1",
):
kind = event["event"]
if kind == "on_chat_model_stream":
print(
f"Chat model chunk: {repr(event['data']['chunk'].content)}",
flush=True,
)
if kind == "on_parser_stream":
print(f"Parser chunk: {event['data']['chunk']}", flush=True)
num_events += 1
if num_events > 30:
# Truncate the output
print("...")
break
Chat model chunk: ' Here'
Chat model chunk: ' is'
Chat model chunk: ' the'
Chat model chunk: ' JSON'
Chat model chunk: ' with'
Chat model chunk: ' the'
Chat model chunk: ' requested'
Chat model chunk: ' countries'
Chat model chunk: ' and'
Chat model chunk: ' their'
Chat model chunk: ' populations'
Chat model chunk: ':'
Chat model chunk: '\n\n```'
Chat model chunk: 'json'
Parser chunk: {}
Chat model chunk: '\n{'
Chat model chunk: '\n '
Chat model chunk: ' "'
Chat model chunk: 'countries'
Chat model chunk: '":'
Parser chunk: {'countries': []}
Chat model chunk: ' ['
Chat model chunk: '\n '
Parser chunk: {'countries': [{}]}
Chat model chunk: ' {'
...
Because both the model and the parser support streaming, we see sreaming events from both components in real time! Kind of cool isn't it? 🦜
Filtering Events
Because this API produces so many events, it is useful to be able to filter on events.
You can filter by either component name
, component tags
or component type
.
By Name
chain = model.with_config({"run_name": "model"}) | JsonOutputParser().with_config(
{"run_name": "my_parser"}
)
max_events = 0
async for event in chain.astream_events(
'output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`',
version="v1",
include_names=["my_parser"],
):
print(event)
max_events += 1
if max_events > 10:
# Truncate output
print("...")
break
{'event': 'on_parser_start', 'name': 'my_parser', 'run_id': 'f2ac1d1c-e14a-45fc-8990-e5c24e707299', 'tags': ['seq:step:2'], 'metadata': {}, 'data': {}}
{'event': 'on_parser_stream', 'name': 'my_parser', 'run_id': 'f2ac1d1c-e14a-45fc-8990-e5c24e707299', 'tags': ['seq:step:2'], 'metadata': {}, 'data': {'chunk': {}}}
{'event': 'on_parser_stream', 'name': 'my_parser', 'run_id': 'f2ac1d1c-e14a-45fc-8990-e5c24e707299', 'tags': ['seq:step:2'], 'metadata': {}, 'data': {'chunk': {'countries': []}}}
{'event': 'on_parser_stream', 'name': 'my_parser', 'run_id': 'f2ac1d1c-e14a-45fc-8990-e5c24e707299', 'tags': ['seq:step:2'], 'metadata': {}, 'data': {'chunk': {'countries': [{}]}}}
{'event': 'on_parser_stream', 'name': 'my_parser', 'run_id': 'f2ac1d1c-e14a-45fc-8990-e5c24e707299', 'tags': ['seq:step:2'], 'metadata': {}, 'data': {'chunk': {'countries': [{'name': ''}]}}}
{'event': 'on_parser_stream', 'name': 'my_parser', 'run_id': 'f2ac1d1c-e14a-45fc-8990-e5c24e707299', 'tags': ['seq:step:2'], 'metadata': {}, 'data': {'chunk': {'countries': [{'name': 'France'}]}}}
{'event': 'on_parser_stream', 'name': 'my_parser', 'run_id': 'f2ac1d1c-e14a-45fc-8990-e5c24e707299', 'tags': ['seq:step:2'], 'metadata': {}, 'data': {'chunk': {'countries': [{'name': 'France', 'population': 67}]}}}
{'event': 'on_parser_stream', 'name': 'my_parser', 'run_id': 'f2ac1d1c-e14a-45fc-8990-e5c24e707299', 'tags': ['seq:step:2'], 'metadata': {}, 'data': {'chunk': {'countries': [{'name': 'France', 'population': 6739}]}}}
{'event': 'on_parser_stream', 'name': 'my_parser', 'run_id': 'f2ac1d1c-e14a-45fc-8990-e5c24e707299', 'tags': ['seq:step:2'], 'metadata': {}, 'data': {'chunk': {'countries': [{'name': 'France', 'population': 673915}]}}}
{'event': 'on_parser_stream', 'name': 'my_parser', 'run_id': 'f2ac1d1c-e14a-45fc-8990-e5c24e707299', 'tags': ['seq:step:2'], 'metadata': {}, 'data': {'chunk': {'countries': [{'name': 'France', 'population': 67391582}]}}}
{'event': 'on_parser_stream', 'name': 'my_parser', 'run_id': 'f2ac1d1c-e14a-45fc-8990-e5c24e707299', 'tags': ['seq:step:2'], 'metadata': {}, 'data': {'chunk': {'countries': [{'name': 'France', 'population': 67391582}, {}]}}}
...
By Type
chain = model.with_config({"run_name": "model"}) | JsonOutputParser().with_config(
{"run_name": "my_parser"}
)
max_events = 0
async for event in chain.astream_events(
'output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`',
version="v1",
include_types=["chat_model"],
):
print(event)
max_events += 1
if max_events > 10:
# Truncate output
print("...")
break
{'event': 'on_chat_model_start', 'name': 'model', 'run_id': '98a6e192-8159-460c-ba73-6dfc921e3777', 'tags': ['seq:step:1'], 'metadata': {}, 'data': {'input': {'messages': [[HumanMessage(content='output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`')]]}}}
{'event': 'on_chat_model_stream', 'name': 'model', 'run_id': '98a6e192-8159-460c-ba73-6dfc921e3777', 'tags': ['seq:step:1'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' Here')}}
{'event': 'on_chat_model_stream', 'name': 'model', 'run_id': '98a6e192-8159-460c-ba73-6dfc921e3777', 'tags': ['seq:step:1'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' is')}}
{'event': 'on_chat_model_stream', 'name': 'model', 'run_id': '98a6e192-8159-460c-ba73-6dfc921e3777', 'tags': ['seq:step:1'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' the')}}
{'event': 'on_chat_model_stream', 'name': 'model', 'run_id': '98a6e192-8159-460c-ba73-6dfc921e3777', 'tags': ['seq:step:1'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' JSON')}}
{'event': 'on_chat_model_stream', 'name': 'model', 'run_id': '98a6e192-8159-460c-ba73-6dfc921e3777', 'tags': ['seq:step:1'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' with')}}
{'event': 'on_chat_model_stream', 'name': 'model', 'run_id': '98a6e192-8159-460c-ba73-6dfc921e3777', 'tags': ['seq:step:1'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' the')}}
{'event': 'on_chat_model_stream', 'name': 'model', 'run_id': '98a6e192-8159-460c-ba73-6dfc921e3777', 'tags': ['seq:step:1'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' requested')}}
{'event': 'on_chat_model_stream', 'name': 'model', 'run_id': '98a6e192-8159-460c-ba73-6dfc921e3777', 'tags': ['seq:step:1'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' countries')}}
{'event': 'on_chat_model_stream', 'name': 'model', 'run_id': '98a6e192-8159-460c-ba73-6dfc921e3777', 'tags': ['seq:step:1'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' and')}}
{'event': 'on_chat_model_stream', 'name': 'model', 'run_id': '98a6e192-8159-460c-ba73-6dfc921e3777', 'tags': ['seq:step:1'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' their')}}
...
By Tags
:::{.callout-caution}
Tags are inherited by child components of a given runnable.
If you're using tags to filter, make sure that this is what you want. :::
chain = (model | JsonOutputParser()).with_config({"tags": ["my_chain"]})
max_events = 0
async for event in chain.astream_events(
'output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`',
version="v1",
include_tags=["my_chain"],
):
print(event)
max_events += 1
if max_events > 10:
# Truncate output
print("...")
break
{'event': 'on_chain_start', 'run_id': '190875f3-3fb7-49ad-9b6e-f49da22f3e49', 'name': 'RunnableSequence', 'tags': ['my_chain'], 'metadata': {}, 'data': {'input': 'output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`'}}
{'event': 'on_chat_model_start', 'name': 'ChatAnthropic', 'run_id': 'ff58f732-b494-4ff9-852a-783d42f4455d', 'tags': ['seq:step:1', 'my_chain'], 'metadata': {}, 'data': {'input': {'messages': [[HumanMessage(content='output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`')]]}}}
{'event': 'on_parser_start', 'name': 'JsonOutputParser', 'run_id': '3b5e4ca1-40fe-4a02-9a19-ba2a43a6115c', 'tags': ['seq:step:2', 'my_chain'], 'metadata': {}, 'data': {}}
{'event': 'on_chat_model_stream', 'name': 'ChatAnthropic', 'run_id': 'ff58f732-b494-4ff9-852a-783d42f4455d', 'tags': ['seq:step:1', 'my_chain'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' Here')}}
{'event': 'on_chat_model_stream', 'name': 'ChatAnthropic', 'run_id': 'ff58f732-b494-4ff9-852a-783d42f4455d', 'tags': ['seq:step:1', 'my_chain'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' is')}}
{'event': 'on_chat_model_stream', 'name': 'ChatAnthropic', 'run_id': 'ff58f732-b494-4ff9-852a-783d42f4455d', 'tags': ['seq:step:1', 'my_chain'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' the')}}
{'event': 'on_chat_model_stream', 'name': 'ChatAnthropic', 'run_id': 'ff58f732-b494-4ff9-852a-783d42f4455d', 'tags': ['seq:step:1', 'my_chain'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' JSON')}}
{'event': 'on_chat_model_stream', 'name': 'ChatAnthropic', 'run_id': 'ff58f732-b494-4ff9-852a-783d42f4455d', 'tags': ['seq:step:1', 'my_chain'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' with')}}
{'event': 'on_chat_model_stream', 'name': 'ChatAnthropic', 'run_id': 'ff58f732-b494-4ff9-852a-783d42f4455d', 'tags': ['seq:step:1', 'my_chain'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' the')}}
{'event': 'on_chat_model_stream', 'name': 'ChatAnthropic', 'run_id': 'ff58f732-b494-4ff9-852a-783d42f4455d', 'tags': ['seq:step:1', 'my_chain'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' requested')}}
{'event': 'on_chat_model_stream', 'name': 'ChatAnthropic', 'run_id': 'ff58f732-b494-4ff9-852a-783d42f4455d', 'tags': ['seq:step:1', 'my_chain'], 'metadata': {}, 'data': {'chunk': AIMessageChunk(content=' countries')}}
...
Non-streaming components
Remember how some components don't stream well because they don't operate on input streams?
While such components can break streaming of the final output when using astream
, astream_events
will still yield streaming events from intermediate steps that support streaming!
# Function that does not support streaming.
# It operates on the finalizes inputs rather than
# operating on the input stream.
def _extract_country_names(inputs):
"""A function that does not operates on input streams and breaks streaming."""
if not isinstance(inputs, dict):
return ""
if "countries" not in inputs:
return ""
countries = inputs["countries"]
if not isinstance(countries, list):
return ""
country_names = [
country.get("name") for country in countries if isinstance(country, dict)
]
return country_names
chain = (
model | JsonOutputParser() | _extract_country_names
) # This parser only works with OpenAI right now
As expected, the astream
API doesn't work correctly because _extract_country_names
doesn't operate on streams.
async for chunk in chain.astream(
'output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`',
):
print(chunk, flush=True)
['France', 'Spain', 'Japan']
Now, let's confirm that with astream_events we're still seeing streaming output from the model and the parser.
num_events = 0
async for event in chain.astream_events(
'output a list of the countries france, spain and japan and their populations in JSON format. Use a dict with an outer key of "countries" which contains a list of countries. Each country should have the key `name` and `population`',
version="v1",
):
kind = event["event"]
if kind == "on_chat_model_stream":
print(
f"Chat model chunk: {repr(event['data']['chunk'].content)}",
flush=True,
)
if kind == "on_parser_stream":
print(f"Parser chunk: {event['data']['chunk']}", flush=True)
num_events += 1
if num_events > 30:
# Truncate the output
print("...")
break
Chat model chunk: ' Here'
Chat model chunk: ' is'
Chat model chunk: ' the'
Chat model chunk: ' JSON'
Chat model chunk: ' with'
Chat model chunk: ' the'
Chat model chunk: ' requested'
Chat model chunk: ' countries'
Chat model chunk: ' and'
Chat model chunk: ' their'
Chat model chunk: ' populations'
Chat model chunk: ':'
Chat model chunk: '\n\n```'
Chat model chunk: 'json'
Parser chunk: {}
Chat model chunk: '\n{'
Chat model chunk: '\n '
Chat model chunk: ' "'
Chat model chunk: 'countries'
Chat model chunk: '":'
Parser chunk: {'countries': []}
Chat model chunk: ' ['
Chat model chunk: '\n '
Parser chunk: {'countries': [{}]}
Chat model chunk: ' {'
Chat model chunk: '\n '
Chat model chunk: ' "'
...
Propagating Callbacks
:::{.callout-caution} If you're using invoking runnables inside your tools, you need to propagate callbacks to the runnable; otherwise, no stream events will be generated. :::
:::{.callout-note} When using RunnableLambdas or @chain decorator, callbacks are propagated automatically behind the scenes. :::
from langchain_core.runnables import RunnableLambda
from langchain_core.tools import tool
def reverse_word(word: str):
return word[::-1]
reverse_word = RunnableLambda(reverse_word)
@tool
def bad_tool(word: str):
"""Custom tool that doesn't propagate callbacks."""
return reverse_word.invoke(word)
async for event in bad_tool.astream_events("hello", version="v1"):
print(event)
{'event': 'on_tool_start', 'run_id': 'ae7690f8-ebc9-4886-9bbe-cb336ff274f2', 'name': 'bad_tool', 'tags': [], 'metadata': {}, 'data': {'input': 'hello'}}
{'event': 'on_tool_stream', 'run_id': 'ae7690f8-ebc9-4886-9bbe-cb336ff274f2', 'tags': [], 'metadata': {}, 'name': 'bad_tool', 'data': {'chunk': 'olleh'}}
{'event': 'on_tool_end', 'name': 'bad_tool', 'run_id': 'ae7690f8-ebc9-4886-9bbe-cb336ff274f2', 'tags': [], 'metadata': {}, 'data': {'output': 'olleh'}}
Here's a re-implementation that does propagate callbacks correctly. You'll notice that now we're getting events from the reverse_word
runnable as well.
@tool
def correct_tool(word: str, callbacks):
"""A tool that correctly propagates callbacks."""
return reverse_word.invoke(word, {"callbacks": callbacks})
async for event in correct_tool.astream_events("hello", version="v1"):
print(event)
{'event': 'on_tool_start', 'run_id': '384f1710-612e-4022-a6d4-8a7bb0cc757e', 'name': 'correct_tool', 'tags': [], 'metadata': {}, 'data': {'input': 'hello'}}
{'event': 'on_chain_start', 'name': 'reverse_word', 'run_id': 'c4882303-8867-4dff-b031-7d9499b39dda', 'tags': [], 'metadata': {}, 'data': {'input': 'hello'}}
{'event': 'on_chain_end', 'name': 'reverse_word', 'run_id': 'c4882303-8867-4dff-b031-7d9499b39dda', 'tags': [], 'metadata': {}, 'data': {'input': 'hello', 'output': 'olleh'}}
{'event': 'on_tool_stream', 'run_id': '384f1710-612e-4022-a6d4-8a7bb0cc757e', 'tags': [], 'metadata': {}, 'name': 'correct_tool', 'data': {'chunk': 'olleh'}}
{'event': 'on_tool_end', 'name': 'correct_tool', 'run_id': '384f1710-612e-4022-a6d4-8a7bb0cc757e', 'tags': [], 'metadata': {}, 'data': {'output': 'olleh'}}
If you're invoking runnables from within Runnable Lambdas or @chains, then callbacks will be passed automatically on your behalf.
from langchain_core.runnables import RunnableLambda
async def reverse_and_double(word: str):
return await reverse_word.ainvoke(word) * 2
reverse_and_double = RunnableLambda(reverse_and_double)
await reverse_and_double.ainvoke("1234")
async for event in reverse_and_double.astream_events("1234", version="v1"):
print(event)
{'event': 'on_chain_start', 'run_id': '4fe56c7b-6982-4999-a42d-79ba56151176', 'name': 'reverse_and_double', 'tags': [], 'metadata': {}, 'data': {'input': '1234'}}
{'event': 'on_chain_start', 'name': 'reverse_word', 'run_id': '335fe781-8944-4464-8d2e-81f61d1f85f5', 'tags': [], 'metadata': {}, 'data': {'input': '1234'}}
{'event': 'on_chain_end', 'name': 'reverse_word', 'run_id': '335fe781-8944-4464-8d2e-81f61d1f85f5', 'tags': [], 'metadata': {}, 'data': {'input': '1234', 'output': '4321'}}
{'event': 'on_chain_stream', 'run_id': '4fe56c7b-6982-4999-a42d-79ba56151176', 'tags': [], 'metadata': {}, 'name': 'reverse_and_double', 'data': {'chunk': '43214321'}}
{'event': 'on_chain_end', 'name': 'reverse_and_double', 'run_id': '4fe56c7b-6982-4999-a42d-79ba56151176', 'tags': [], 'metadata': {}, 'data': {'output': '43214321'}}
And with the @chain decorator:
from langchain_core.runnables import chain
@chain
async def reverse_and_double(word: str):
return await reverse_word.ainvoke(word) * 2
await reverse_and_double.ainvoke("1234")
async for event in reverse_and_double.astream_events("1234", version="v1"):
print(event)
{'event': 'on_chain_start', 'run_id': '7485eedb-1854-429c-a2f8-03d01452daef', 'name': 'reverse_and_double', 'tags': [], 'metadata': {}, 'data': {'input': '1234'}}
{'event': 'on_chain_start', 'name': 'reverse_word', 'run_id': 'e7cddab2-9b95-4e80-abaf-4b2429117835', 'tags': [], 'metadata': {}, 'data': {'input': '1234'}}
{'event': 'on_chain_end', 'name': 'reverse_word', 'run_id': 'e7cddab2-9b95-4e80-abaf-4b2429117835', 'tags': [], 'metadata': {}, 'data': {'input': '1234', 'output': '4321'}}
{'event': 'on_chain_stream', 'run_id': '7485eedb-1854-429c-a2f8-03d01452daef', 'tags': [], 'metadata': {}, 'name': 'reverse_and_double', 'data': {'chunk': '43214321'}}
{'event': 'on_chain_end', 'name': 'reverse_and_double', 'run_id': '7485eedb-1854-429c-a2f8-03d01452daef', 'tags': [], 'metadata': {}, 'data': {'output': '43214321'}}