- Permit the specification of a `root_dir` to the read/write file tools
to specify a working directory
- Add validation for attempts to read/write outside the directory (e.g.,
through `../../` or symlinks or `/abs/path`'s that don't lie in the
correct path)
- Add some tests for all
One question is whether we should make a default root directory for
these? tradeoffs either way
`langchain.prompts.PromptTemplate` and
`langchain.prompts.FewShotPromptTemplate` do not validate
`input_variables` when initialized as `jinja2` template.
```python
# Using langchain v0.0.144
template = """"\
Your variable: {{ foo }}
{% if bar %}
You just set bar boolean variable to true
{% endif %}
"""
# Missing variable, should raise ValueError
prompt_template = PromptTemplate(template=template,
input_variables=["bar"],
template_format="jinja2",
validate_template=True)
# Extra variable, should raise ValueError
prompt_template = PromptTemplate(template=template,
input_variables=["bar", "foo", "extra", "thing"],
template_format="jinja2",
validate_template=True)
```
- Remove dynamic model creation in the `args()` property. _Only infer
for the decorator (and add an argument to NOT infer if someone wishes to
only pass as a string)_
- Update the validation example to make it less likely to be
misinterpreted as a "safe" way to run a repl
There is one example of "Multi-argument tools" in the custom_tools.ipynb
from yesterday, but we could add more. The output parsing for the base
MRKL agent hasn't been adapted to handle structured args at this point
in time
---------
Co-authored-by: Harrison Chase <hw.chase.17@gmail.com>
`langchain.prompts.PromptTemplate` is unable to infer `input_variables`
from jinja2 template.
```python
# Using langchain v0.0.141
template_string = """\
Hello world
Your variable: {{ var }}
{# This will not get rendered #}
{% if verbose %}
Congrats! You just turned on verbose mode and got extra messages!
{% endif %}
"""
template = PromptTemplate.from_template(template_string, template_format="jinja2")
print(template.input_variables) # Output ['# This will not get rendered #', '% endif %', '% if verbose %']
```
---------
Co-authored-by: engkheng <ongengkheng929@example.com>
Add a time-weighted memory retriever and a notebook that approximates a
Generative Agent from https://arxiv.org/pdf/2304.03442.pdf
The "daily plan" components are removed for now since they are less
useful without a virtual world, but the memory is an interesting
component to build off.
---------
Co-authored-by: Harrison Chase <hw.chase.17@gmail.com>
Use numexpr evaluate instead of the python REPL to avoid malicious code
injection.
Tested against the (limited) math dataset and got the same score as
before.
For more permissive tools (like the REPL tool itself), other approaches
ought to be provided (some combination of Sanitizer + Restricted python
+ unprivileged-docker + ...), but for a calculator tool, only
mathematical expressions should be permitted.
See https://github.com/hwchase17/langchain/issues/814
When the code ran by the PythonAstREPLTool contains multiple statements
it will fallback to exec() instead of using eval(). With this change, it
will also return the output of the code in the same way the
PythonREPLTool will.
Currently, the output type of a number of OutputParser's `parse` methods
is `Any` when it can in fact be inferred.
This PR makes BaseOutputParser use a generic type and fixes the output
types of the following parsers:
- `PydanticOutputParser`
- `OutputFixingParser`
- `RetryOutputParser`
- `RetryWithErrorOutputParser`
The output of the `StructuredOutputParser` is corrected from `BaseModel`
to `Any` since there are no type guarantees provided by the parser.
Fixes issue #2715
`combine_docs` does not go through the standard chain call path which
means that chain callbacks won't be triggered, meaning QA chains won't
be traced properly, this fixes that.
Also fix several errors in the chat_vector_db notebook
Right now, eval chains require an answer for every question. It's
cumbersome to collect this ground truth so getting around this issue
with 2 things:
* Adding a context param in `ContextQAEvalChain` and simply evaluating
if the question is answered accurately from context
* Adding chain of though explanation prompting to improve the accuracy
of this w/o GT.
This also gets to feature parity with openai/evals which has the same
contextual eval w/o GT.
TODO in follow-up:
* Better prompt inheritance. No need for seperate prompt for CoT
reasoning. How can we merge them together
---------
Co-authored-by: Vashisht Madhavan <vashishtmadhavan@Vashs-MacBook-Pro.local>
This still doesn't handle the following
- non-JSON media types
- anyOf, allOf, oneOf's
And doesn't emit the typescript definitions for referred types yet, but
that can be saved for a separate PR.
Also, we could have better support for Swagger 2.0 specs and OpenAPI
3.0.3 (can use the same lib for the latter) recommend offline conversion
for now.
`AgentExecutor` already has support for limiting the number of
iterations. But the amount of time taken for each iteration can vary
quite a bit, so it is difficult to place limits on the execution time.
This PR adds a new field `max_execution_time` to the `AgentExecutor`
model. When called asynchronously, the agent loop is wrapped in an
`asyncio.timeout()` context which triggers the early stopping response
if the time limit is reached. When called synchronously, the agent loop
checks for both the max_iteration limit and the time limit after each
iteration.
When used asynchronously `max_execution_time` gives really tight control
over the max time for an execution chain. When used synchronously, the
chain can unfortunately exceed max_execution_time, but it still gives
more control than trying to estimate the number of max_iterations needed
to cap the execution time.
---------
Co-authored-by: Zachary Jones <zjones@zetaglobal.com>
This pull request adds an enum class for the various types of agents
used in the project, located in the `agent_types.py` file. Currently,
the project is using hardcoded strings for the initialization of these
agents, which can lead to errors and make the code harder to maintain.
With the introduction of the new enums, the code will be more readable
and less error-prone.
The new enum members include:
- ZERO_SHOT_REACT_DESCRIPTION
- REACT_DOCSTORE
- SELF_ASK_WITH_SEARCH
- CONVERSATIONAL_REACT_DESCRIPTION
- CHAT_ZERO_SHOT_REACT_DESCRIPTION
- CHAT_CONVERSATIONAL_REACT_DESCRIPTION
In this PR, I have also replaced the hardcoded strings with the
appropriate enum members throughout the codebase, ensuring a smooth
transition to the new approach.
Fix issue#1645: Parse either whitespace or newline after 'Action Input:'
in llm_output in mrkl agent.
Unittests added accordingly.
Co-authored-by: ₿ingnan.ΞTH <brillliantz@outlook.com>
I was getting the same issue reported in #1339 by
[MacYang555](https://github.com/MacYang555) when running the test suite
on my Mac. I implemented the fix they suggested to use a regex match in
the output assertion for the scenario under test.
Resolves#1339
Seeing a lot of issues in Discord in which the LLM is not using the
correct LIMIT clause for different SQL dialects. ie, it's using `LIMIT`
for mssql instead of `TOP`, or instead of `ROWNUM` for Oracle, etc.
I think this could be due to us specifying the LIMIT statement in the
example rows portion of `table_info`. So the LLM is seeing the `LIMIT`
statement used in the prompt.
Since we can't specify each dialect's method here, I think it's fine to
just replace the `SELECT... LIMIT 3;` statement with `3 rows from
table_name table:`, and wrap everything in a block comment directly
following the `CREATE` statement. The Rajkumar et al paper wrapped the
example rows and `SELECT` statement in a block comment as well anyway.
Thoughts @fpingham?
This class enables us to send a dictionary containing an output key and
the expected format, which in turn allows us to retrieve the result of
the matching formats and extract specific information from it.
To exclude irrelevant information from our return dictionary, we can
prompt the LLM to use a specific command that notifies us when it
doesn't know the answer. We refer to this variable as the
"no_update_value".
Regarding the updated regular expression pattern
(r"{}:\s?([^.'\n']*).?"), it enables us to retrieve a format as 'Output
Key':'value'.
We have improved the regex by adding an optional space between ':' and
'value' with "s?", and by excluding points and line jumps from the
matches using "[^.'\n']*".
Provide shared memory capability for the Agent.
Inspired by #1293 .
## Problem
If both Agent and Tools (i.e., LLMChain) use the same memory, both of
them will save the context. It can be annoying in some cases.
## Solution
Create a memory wrapper that ignores the save and clear, thereby
preventing updates from Agent or Tools.
This PR adds
* `ZeroShotAgent.as_sql_agent`, which returns an agent for interacting
with a sql database. This builds off of `SQLDatabaseChain`. The main
advantages are 1) answering general questions about the db, 2) access to
a tool for double checking queries, and 3) recovering from errors
* `ZeroShotAgent.as_json_agent` which returns an agent for interacting
with json blobs.
* Several examples in notebooks
---------
Co-authored-by: Harrison Chase <hw.chase.17@gmail.com>
Follow-up of @hinthornw's PR:
- Migrate the Tool abstraction to a separate file (`BaseTool`).
- `Tool` implementation of `BaseTool` takes in function and coroutine to
more easily maintain backwards compatibility
- Add a Toolkit abstraction that can own the generation of tools around
a shared concept or state
---------
Co-authored-by: William FH <13333726+hinthornw@users.noreply.github.com>
Co-authored-by: Harrison Chase <hw.chase.17@gmail.com>
Co-authored-by: Francisco Ingham <fpingham@gmail.com>
Co-authored-by: Dhruv Anand <105786647+dhruv-anand-aintech@users.noreply.github.com>
Co-authored-by: cragwolfe <cragcw@gmail.com>
Co-authored-by: Anton Troynikov <atroyn@users.noreply.github.com>
Co-authored-by: Oliver Klingefjord <oliver@klingefjord.com>
Co-authored-by: William Fu-Hinthorn <whinthorn@Williams-MBP-3.attlocal.net>
Co-authored-by: Bruno Bornsztein <bruno.bornsztein@gmail.com>
This approach has several advantages:
* it improves the readability of the code
* removes incompatibilities between SQL dialects
* fixes a bug with `datetime` values in rows and `ast.literal_eval`
Huge thanks and credits to @jzluo for finding the weaknesses in the
current approach and for the thoughtful discussion on the best way to
implement this.
---------
Co-authored-by: Francisco Ingham <>
Co-authored-by: Jon Luo <20971593+jzluo@users.noreply.github.com>
Currently the chain is getting the column names and types on the one
side and the example rows on the other. It is easier for the llm to read
the table information if the column name and examples are shown together
so that it can easily understand to which columns do the examples refer
to. For an instantiation of this, please refer to the changes in the
`sqlite.ipynb` notebook.
Also changed `eval` for `ast.literal_eval` when interpreting the results
from the sample row query since it is a better practice.
---------
Co-authored-by: Francisco Ingham <>
---------
Co-authored-by: Francisco Ingham <fpingham@gmail.com>
Supporting asyncio in langchain primitives allows for users to run them
concurrently and creates more seamless integration with
asyncio-supported frameworks (FastAPI, etc.)
Summary of changes:
**LLM**
* Add `agenerate` and `_agenerate`
* Implement in OpenAI by leveraging `client.Completions.acreate`
**Chain**
* Add `arun`, `acall`, `_acall`
* Implement them in `LLMChain` and `LLMMathChain` for now
**Agent**
* Refactor and leverage async chain and llm methods
* Add ability for `Tools` to contain async coroutine
* Implement async SerpaPI `arun`
Create demo notebook.
Open questions:
* Should all the async stuff go in separate classes? I've seen both
patterns (keeping the same class and having async and sync methods vs.
having class separation)
This allows the LLM to correct its previous command by looking at the
error message output to the shell.
Additionally, this uses subprocess.run because that is now recommended
over subprocess.check_output:
https://docs.python.org/3/library/subprocess.html#using-the-subprocess-module
Co-authored-by: Amos Ng <me@amos.ng>
# Problem
I noticed that in order to change the prefix of the prompt in the
`zero-shot-react-description` agent
we had to dig around to subset strings deep into the agent's attributes.
It requires the user to inspect a long chain of attributes and classes.
`initialize_agent -> AgentExecutor -> Agent -> LLMChain -> Prompt from
Agent.create_prompt`
``` python
agent = initialize_agent(
tools=tools,
llm=fake_llm,
agent="zero-shot-react-description"
)
prompt_str = agent.agent.llm_chain.prompt.template
new_prompt_str = change_prefix(prompt_str)
agent.agent.llm_chain.prompt.template = new_prompt_str
```
# Implemented Solution
`initialize_agent` accepts `**kwargs` but passes it to `AgentExecutor`
but not `ZeroShotAgent`, by simply giving the kwargs to the agent class
methods we can support changing the prefix and suffix for one agent
while allowing future agents to take advantage of `initialize_agent`.
```
agent = initialize_agent(
tools=tools,
llm=fake_llm,
agent="zero-shot-react-description",
agent_kwargs={"prefix": prefix, "suffix": suffix}
)
```
To be fair, this was before finding docs around custom agents here:
https://langchain.readthedocs.io/en/latest/modules/agents/examples/custom_agent.html?highlight=custom%20#custom-llmchain
but i find that my use case just needed to change the prefix a little.
# Changes
* Pass kwargs to Agent class method
* Added a test to check suffix and prefix
---------
Co-authored-by: Jason Liu <jason@jxnl.coA>
It's generally considered to be a good practice to pin dependencies to
prevent surprise breakages when a new version of a dependency is
released. This commit adds the ability to pin dependencies when loading
from LangChainHub.
Centralizing this logic and using urllib fixes an issue identified by
some windows users highlighted in this video -
https://youtu.be/aJ6IQUh8MLQ?t=537
The agents usually benefit from understanding what the data looks like
to be able to filter effectively. Sending just one row in the table info
allows the agent to understand the data before querying and get better
results.
---------
Co-authored-by: Francisco Ingham <>
---------
Co-authored-by: Francisco Ingham <fpingham@gmail.com>
* add implementations of `BaseCallbackHandler` to support tracing:
`SharedTracer` which is thread-safe and `Tracer` which is not and is
meant to be used locally.
* Tracers persist runs to locally running `langchain-server`
Co-authored-by: Harrison Chase <hw.chase.17@gmail.com>
Big docs refactor! Motivation is to make it easier for people to find
resources they are looking for. To accomplish this, there are now three
main sections:
- Getting Started: steps for getting started, walking through most core
functionality
- Modules: these are different modules of functionality that langchain
provides. Each part here has a "getting started", "how to", "key
concepts" and "reference" section (except in a few select cases where it
didnt easily fit).
- Use Cases: this is to separate use cases (like summarization, question
answering, evaluation, etc) from the modules, and provide a different
entry point to the code base.
There is also a full reference section, as well as extra resources
(glossary, gallery, etc)
Co-authored-by: Shreya Rajpal <ShreyaR@users.noreply.github.com>
This PR has two contributions:
1. Add test for when stop token is found in middle of text
2. Add code coverage tooling and instructions
- Add pytest-cov via poetry
- Add necessary config files
- Add new make instruction for `coverage`
- Update README with coverage guidance
- Update minor README formatting/spelling
Co-authored-by: Hunter Gerlach <hunter@huntergerlach.com>
Love the project, a ton of fun!
I think the PR is pretty self-explanatory, happy to make any changes! I
am working on using it in an `LLMBashChain` and may update as that
progresses.
Co-authored-by: Harrison Chase <hw.chase.17@gmail.com>
Add support for calling HuggingFace embedding models
using the HuggingFaceHub Inference API. New class mirrors
the existing HuggingFaceHub LLM implementation. Currently
only supports 'sentence-transformers' models.
Closes#86
Add MemoryChain and ConversationChain as chains that take a docstore in
addition to the prompt, and use the docstore to stuff context into the
prompt. This can be used to have an ongoing conversation with a chatbot.
Probably needs a bit of refactoring for code quality
Co-authored-by: Harrison Chase <hw.chase.17@gmail.com>
Also updated docs, and noticed an issue with the add_texts method on
VectorStores that I had missed before -- the metadatas arg should be
required to match the classmethod which initializes the VectorStores
(the add_example methods break otherwise in the ExampleSelectors)
this will break atm but wanted to get thoughts on implementation.
1. should add() be on docstore interface?
2. should InMemoryDocstore change to take a list of documents as init?
(makes this slightly easier to implement in FAISS -- if we think it is
less clean then could expose a method to get the number of documents
currently in the dict, and perform the logic of creating the necessary
dictionary in the FAISS.add_texts method.
Co-authored-by: Harrison Chase <hw.chase.17@gmail.com>
`SQLDatabase` now accepts two `init` arguments:
1. `ignore_tables` to pass in a list of tables to not search over
2. `include_tables` to restrict to a list of tables to consider
This is a simple proof of concept of using external files as templates.
I'm still feeling my way around the codebase.
As a user, I want to use files as prompts, so it will be easier to
manage and test prompts.
The future direction is to use a template engine, most likely Mako.
