Adds a call to Pydantic's `update_forward_refs` for the `Run` class (in
addition to the `ChainRun` and `ToolRun` classes, for which that method
is already called). Without it, the self-reference of child classes
(type `List[Run]`) is problematic. For example:
```python
from langchain.callbacks import StdOutCallbackHandler
from langchain.chains import LLMChain
from langchain.llms import OpenAI
from langchain.prompts import PromptTemplate
from wandb.integration.langchain import WandbTracer
llm = OpenAI()
prompt = PromptTemplate.from_template("1 + {number} = ")
chain = LLMChain(llm=llm, prompt=prompt, callbacks=[StdOutCallbackHandler(), WandbTracer()])
print(chain.run(number=2))
```
results in the following output before the change
```
WARNING:root:Error in on_chain_start callback: field "child_runs" not yet prepared so type is still a ForwardRef, you might need to call Run.update_forward_refs().
> Entering new LLMChain chain...
Prompt after formatting:
1 + 2 =
WARNING:root:Error in on_chain_end callback: No chain Run found to be traced
> Finished chain.
3
```
but afterwards the callback error messages are gone.
Hi there!
I'm excited to open this PR to add support for using 'Tencent Cloud
VectorDB' as a vector store.
Tencent Cloud VectorDB is a fully-managed, self-developed,
enterprise-level distributed database service designed for storing,
retrieving, and analyzing multi-dimensional vector data. The database
supports multiple index types and similarity calculation methods, with a
single index supporting vector scales up to 1 billion and capable of
handling millions of QPS with millisecond-level query latency. Tencent
Cloud VectorDB not only provides external knowledge bases for large
models to improve their accuracy, but also has wide applications in AI
fields such as recommendation systems, NLP services, computer vision,
and intelligent customer service.
The PR includes:
Implementation of Vectorstore.
I have read your [contributing
guidelines](72b7d76d79/.github/CONTRIBUTING.md).
And I have passed the tests below
make format
make lint
make coverage
make test
This PR brings structural updates to `PlaywrightURLLoader`, aiming at
making the code more readable and extensible through the abstraction of
page evaluation logic. These changes also align this implementation with
a similar structure used in LangChain.js.
The key enhancements include:
1. Introduction of 'PlaywrightEvaluator', an abstract base class for all
evaluators.
2. Creation of 'UnstructuredHtmlEvaluator', a concrete class
implementing 'PlaywrightEvaluator', which uses `unstructured` library
for processing page's HTML content.
3. Extension of 'PlaywrightURLLoader' constructor to optionally accept
an evaluator of the type 'PlaywrightEvaluator'. It defaults to
'UnstructuredHtmlEvaluator' if no evaluator is provided.
4. Refactoring of 'load' and 'aload' methods to use the 'evaluate' and
'evaluate_async' methods of the provided 'PageEvaluator' for page
content handling.
This update brings flexibility to 'PlaywrightURLLoader' as it can now
utilize different evaluators for page processing depending on the
requirement. The abstraction also improves code maintainability and
readability.
Twitter: @ywkim
- Description: Add bloomz_7b, llama-2-7b, llama-2-13b, llama-2-70b to
ErnieBotChat, which only supported ERNIE-Bot-turbo and ERNIE-Bot.
- Issue: #10022,
- Dependencies: no extra dependencies
---------
Co-authored-by: hetianfeng <hetianfeng@meituan.com>
### Description
The feature for anonymizing data has been implemented. In order to
protect private data, such as when querying external APIs (OpenAI), it
is worth pseudonymizing sensitive data to maintain full privacy.
Anonynization consists of two steps:
1. **Identification:** Identify all data fields that contain personally
identifiable information (PII).
2. **Replacement**: Replace all PIIs with pseudo values or codes that do
not reveal any personal information about the individual but can be used
for reference. We're not using regular encryption, because the language
model won't be able to understand the meaning or context of the
encrypted data.
We use *Microsoft Presidio* together with *Faker* framework for
anonymization purposes because of the wide range of functionalities they
provide. The full implementation is available in `PresidioAnonymizer`.
### Future works
- **deanonymization** - add the ability to reverse anonymization. For
example, the workflow could look like this: `anonymize -> LLMChain ->
deanonymize`. By doing this, we will retain anonymity in requests to,
for example, OpenAI, and then be able restore the original data.
- **instance anonymization** - at this point, each occurrence of PII is
treated as a separate entity and separately anonymized. Therefore, two
occurrences of the name John Doe in the text will be changed to two
different names. It is therefore worth introducing support for full
instance detection, so that repeated occurrences are treated as a single
object.
### Twitter handle
@deepsense_ai / @MaksOpp
---------
Co-authored-by: MaksOpp <maks.operlejn@gmail.com>
Co-authored-by: Bagatur <baskaryan@gmail.com>
- Description: this PR adds `s3_object_key` and `s3_bucket` to the doc
metadata when loading an S3 file. This is particularly useful when using
`S3DirectoryLoader` to remove the files from the dir once they have been
processed (getting the object keys from the metadata `source` field
seems brittle)
- Dependencies: N/A
- Tag maintainer: ?
- Twitter handle: _cbornet
---------
Co-authored-by: Eugene Yurtsev <eyurtsev@gmail.com>
This PR makes the following changes:
1. Documents become serializable using langhchain serialization
2. Make a utility to create a docstore kw store
Will help to address issue here:
https://github.com/langchain-ai/langchain/issues/9345
In the function _load_run_evaluators the function _get_keys was not
called if only custom_evaluators parameter is used
- Description: In the function _load_run_evaluators the function
_get_keys was not called if only custom_evaluators parameter is used,
- Issue: no issue created for this yet,
- Dependencies: None,
- Tag maintainer: @vowelparrot,
- Twitter handle: Buckler89
---------
Co-authored-by: ddroghini <d.droghini@mflgroup.com>
Description: This commit uses the new Service object in Selenium
webdriver as executable_path has been [deprecated and removed in
selenium version
4.11.2](9f5801c82f)
Issue: https://github.com/langchain-ai/langchain/issues/9808
Tag Maintainer: @eyurtsev
- Description: In my previous PR, I had modified the code to catch all
kinds of [SOURCES, sources, Source, Sources]. However, this change
included checking for a colon or a white space which should actually
have been only checking for a colon.
- Issue: the issue # it fixes (if applicable),
- Dependencies: any dependencies required for this change,
Adds support for [llmonitor](https://llmonitor.com) callbacks.
It enables:
- Requests tracking / logging / analytics
- Error debugging
- Cost analytics
- User tracking
Let me know if anythings neds to be changed for merge.
Thank you!
Co-authored-by: Daniel Brenot <dbrenot@pelmorex.com>
Co-authored-by: Daniel <daniel.alexander.brenot@gmail.com>
Co-authored-by: Bagatur <baskaryan@gmail.com>
- Description: the implementation for similarity_search_with_score did
not actually include a score or logic to filter. Now fixed.
- Tag maintainer: @rlancemartin
- Twitter handle: @ofermend
Recently we made the decision that PromptGuard takes a list of strings
instead of a string.
@ggroode implemented the integration change.
---------
Co-authored-by: ggroode <ggroode@berkeley.edu>
Co-authored-by: ggroode <46691276+ggroode@users.noreply.github.com>
Clearly document that the PAL and CPAL techniques involve generating
code, and that such code must be properly sandboxed and given
appropriate narrowly-scoped credentials in order to ensure security.
While our implementations include some mitigations, Python and SQL
sandboxing is well-known to be a very hard problem and our mitigations
are no replacement for proper sandboxing and permissions management. The
implementation of such techniques must be performed outside the scope of
the Python process where this package's code runs, so its correct setup
and administration must therefore be the responsibility of the user of
this code.
- Description: added the _cosine_relevance_score_fn to
_select_relevance_score_fn of faiss.py to enable the use of cosine
distance for similarity for this vector store and to comply with the
Error Message, that implies, that cosine should be a valid distance
strategy
- Issue: no relevant Issue found, but needed this function myself and
tested it in a private repo
- Dependencies: none
Neo4j has added vector index integration just recently. To allow both
ingestion and integrating it as vector RAG applications, I wrapped it as
a vector store as the implementation is completely different from
`GraphCypherQAChain`. Here, we are not generating any Cypher statements
at query time, we are simply doing the vector similarity search using
the new vector index as if we were dealing with a vector database.
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
Mypy was not able to determine a good type for `type_to_loader_dict`,
since the values in the dict are functions whose return types are
related to each other in a complex way. One can see this by adding a
line like `reveal_type(type_to_loader_dict)` and running mypy, which
will get mypy to show what type it has inferred for that value.
Adding an explicit type hint to help out mypy avoids the need for a mypy
suppression and allows the code to type-check cleanly.
In order to use `requires` marker in langchain-experimental, there's a
need for *conftest.py* file inside. Everything is identical to the main
langchain module.
Co-authored-by: maks-operlejn-ds <maks.operlejn@gmail.com>
We always overwrote the required args but we infer them by default.
Doing it only the old way makes it so the llm guesses even if an arg is
optional (e.g., for uuids)
The most reliable way to not have a chain run an undesirable SQL command
is to not give it database permissions to run that command. That way the
database itself performs the rule enforcement, so it's much easier to
configure and use properly than anything we could add in ourselves.
## Description
The following PR enables the [grammar-based
sampling](https://github.com/ggerganov/llama.cpp/tree/master/grammars)
in llama-cpp LLM.
In short, loading file with formal grammar definition will constrain
model outputs. For instance, one can force the model to generate valid
JSON or generate only python lists.
In the follow-up PR we will add:
* docs with some description why it is cool and how it works
* maybe some code sample for some task such as in llama repo
---------
Co-authored-by: Lance Martin <lance@langchain.dev>
Co-authored-by: Bagatur <baskaryan@gmail.com>
Expose classmethods to convenient initialize the vectostore.
The purpose of this PR is to make it easy for users to initialize an
empty vectorstore that's properly pre-configured without having to index
documents into it via `from_documents`.
This will make it easier for users to rely on the following indexing
code: https://github.com/langchain-ai/langchain/pull/9614
to help manage data in the qdrant vectorstore.
### Description
The previous Redis implementation did not allow for the user to specify
the index configuration (i.e. changing the underlying algorithm) or add
additional metadata to use for querying (i.e. hybrid or "filtered"
search).
This PR introduces the ability to specify custom index attributes and
metadata attributes as well as use that metadata in filtered queries.
Overall, more structure was introduced to the Redis implementation that
should allow for easier maintainability moving forward.
# New Features
The following features are now available with the Redis integration into
Langchain
## Index schema generation
The schema for the index will now be automatically generated if not
specified by the user. For example, the data above has the multiple
metadata categories. The the following example
```python
from langchain.embeddings import OpenAIEmbeddings
from langchain.vectorstores.redis import Redis
embeddings = OpenAIEmbeddings()
rds, keys = Redis.from_texts_return_keys(
texts,
embeddings,
metadatas=metadata,
redis_url="redis://localhost:6379",
index_name="users"
)
```
Loading the data in through this and the other ``from_documents`` and
``from_texts`` methods will now generate index schema in Redis like the
following.
view index schema with the ``redisvl`` tool. [link](redisvl.com)
```bash
$ rvl index info -i users
```
Index Information:
| Index Name | Storage Type | Prefixes | Index Options | Indexing |
|--------------|----------------|---------------|-----------------|------------|
| users | HASH | ['doc:users'] | [] | 0 |
Index Fields:
| Name | Attribute | Type | Field Option | Option Value |
|----------------|----------------|---------|----------------|----------------|
| user | user | TEXT | WEIGHT | 1 |
| job | job | TEXT | WEIGHT | 1 |
| credit_score | credit_score | TEXT | WEIGHT | 1 |
| content | content | TEXT | WEIGHT | 1 |
| age | age | NUMERIC | | |
| content_vector | content_vector | VECTOR | | |
### Custom Metadata specification
The metadata schema generation has the following rules
1. All text fields are indexed as text fields.
2. All numeric fields are index as numeric fields.
If you would like to have a text field as a tag field, users can specify
overrides like the following for the example data
```python
# this can also be a path to a yaml file
index_schema = {
"text": [{"name": "user"}, {"name": "job"}],
"tag": [{"name": "credit_score"}],
"numeric": [{"name": "age"}],
}
rds, keys = Redis.from_texts_return_keys(
texts,
embeddings,
metadatas=metadata,
redis_url="redis://localhost:6379",
index_name="users"
)
```
This will change the index specification to
Index Information:
| Index Name | Storage Type | Prefixes | Index Options | Indexing |
|--------------|----------------|----------------|-----------------|------------|
| users2 | HASH | ['doc:users2'] | [] | 0 |
Index Fields:
| Name | Attribute | Type | Field Option | Option Value |
|----------------|----------------|---------|----------------|----------------|
| user | user | TEXT | WEIGHT | 1 |
| job | job | TEXT | WEIGHT | 1 |
| content | content | TEXT | WEIGHT | 1 |
| credit_score | credit_score | TAG | SEPARATOR | , |
| age | age | NUMERIC | | |
| content_vector | content_vector | VECTOR | | |
and throw a warning to the user (log output) that the generated schema
does not match the specified schema.
```text
index_schema does not match generated schema from metadata.
index_schema: {'text': [{'name': 'user'}, {'name': 'job'}], 'tag': [{'name': 'credit_score'}], 'numeric': [{'name': 'age'}]}
generated_schema: {'text': [{'name': 'user'}, {'name': 'job'}, {'name': 'credit_score'}], 'numeric': [{'name': 'age'}]}
```
As long as this is on purpose, this is fine.
The schema can be defined as a yaml file or a dictionary
```yaml
text:
- name: user
- name: job
tag:
- name: credit_score
numeric:
- name: age
```
and you pass in a path like
```python
rds, keys = Redis.from_texts_return_keys(
texts,
embeddings,
metadatas=metadata,
redis_url="redis://localhost:6379",
index_name="users3",
index_schema=Path("sample1.yml").resolve()
)
```
Which will create the same schema as defined in the dictionary example
Index Information:
| Index Name | Storage Type | Prefixes | Index Options | Indexing |
|--------------|----------------|----------------|-----------------|------------|
| users3 | HASH | ['doc:users3'] | [] | 0 |
Index Fields:
| Name | Attribute | Type | Field Option | Option Value |
|----------------|----------------|---------|----------------|----------------|
| user | user | TEXT | WEIGHT | 1 |
| job | job | TEXT | WEIGHT | 1 |
| content | content | TEXT | WEIGHT | 1 |
| credit_score | credit_score | TAG | SEPARATOR | , |
| age | age | NUMERIC | | |
| content_vector | content_vector | VECTOR | | |
### Custom Vector Indexing Schema
Users with large use cases may want to change how they formulate the
vector index created by Langchain
To utilize all the features of Redis for vector database use cases like
this, you can now do the following to pass in index attribute modifiers
like changing the indexing algorithm to HNSW.
```python
vector_schema = {
"algorithm": "HNSW"
}
rds, keys = Redis.from_texts_return_keys(
texts,
embeddings,
metadatas=metadata,
redis_url="redis://localhost:6379",
index_name="users3",
vector_schema=vector_schema
)
```
A more complex example may look like
```python
vector_schema = {
"algorithm": "HNSW",
"ef_construction": 200,
"ef_runtime": 20
}
rds, keys = Redis.from_texts_return_keys(
texts,
embeddings,
metadatas=metadata,
redis_url="redis://localhost:6379",
index_name="users3",
vector_schema=vector_schema
)
```
All names correspond to the arguments you would set if using Redis-py or
RedisVL. (put in doc link later)
### Better Querying
Both vector queries and Range (limit) queries are now available and
metadata is returned by default. The outputs are shown.
```python
>>> query = "foo"
>>> results = rds.similarity_search(query, k=1)
>>> print(results)
[Document(page_content='foo', metadata={'user': 'derrick', 'job': 'doctor', 'credit_score': 'low', 'age': '14', 'id': 'doc:users:657a47d7db8b447e88598b83da879b9d', 'score': '7.15255737305e-07'})]
>>> results = rds.similarity_search_with_score(query, k=1, return_metadata=False)
>>> print(results) # no metadata, but with scores
[(Document(page_content='foo', metadata={}), 7.15255737305e-07)]
>>> results = rds.similarity_search_limit_score(query, k=6, score_threshold=0.0001)
>>> print(len(results)) # range query (only above threshold even if k is higher)
4
```
### Custom metadata filtering
A big advantage of Redis in this space is being able to do filtering on
data stored alongside the vector itself. With the example above, the
following is now possible in langchain. The equivalence operators are
overridden to describe a new expression language that mimic that of
[redisvl](redisvl.com). This allows for arbitrarily long sequences of
filters that resemble SQL commands that can be used directly with vector
queries and range queries.
There are two interfaces by which to do so and both are shown.
```python
>>> from langchain.vectorstores.redis import RedisFilter, RedisNum, RedisText
>>> age_filter = RedisFilter.num("age") > 18
>>> age_filter = RedisNum("age") > 18 # equivalent
>>> results = rds.similarity_search(query, filter=age_filter)
>>> print(len(results))
3
>>> job_filter = RedisFilter.text("job") == "engineer"
>>> job_filter = RedisText("job") == "engineer" # equivalent
>>> results = rds.similarity_search(query, filter=job_filter)
>>> print(len(results))
2
# fuzzy match text search
>>> job_filter = RedisFilter.text("job") % "eng*"
>>> results = rds.similarity_search(query, filter=job_filter)
>>> print(len(results))
2
# combined filters (AND)
>>> combined = age_filter & job_filter
>>> results = rds.similarity_search(query, filter=combined)
>>> print(len(results))
1
# combined filters (OR)
>>> combined = age_filter | job_filter
>>> results = rds.similarity_search(query, filter=combined)
>>> print(len(results))
4
```
All the above filter results can be checked against the data above.
### Other
- Issue: #3967
- Dependencies: No added dependencies
- Tag maintainer: @hwchase17 @baskaryan @rlancemartin
- Twitter handle: @sampartee
---------
Co-authored-by: Naresh Rangan <naresh.rangan0@walmart.com>
Co-authored-by: Bagatur <baskaryan@gmail.com>
This PR implements a custom chain that wraps Amazon Comprehend API
calls. The custom chain is aimed to be used with LLM chains to provide
moderation capability that let’s you detect and redact PII, Toxic and
Intent content in the LLM prompt, or the LLM response. The
implementation accepts a configuration object to control what checks
will be performed on a LLM prompt and can be used in a variety of setups
using the LangChain expression language to not only detect the
configured info in chains, but also other constructs such as a
retriever.
The included sample notebook goes over the different configuration
options and how to use it with other chains.
### Usage sample
```python
from langchain_experimental.comprehend_moderation import BaseModerationActions, BaseModerationFilters
moderation_config = {
"filters":[
BaseModerationFilters.PII,
BaseModerationFilters.TOXICITY,
BaseModerationFilters.INTENT
],
"pii":{
"action": BaseModerationActions.ALLOW,
"threshold":0.5,
"labels":["SSN"],
"mask_character": "X"
},
"toxicity":{
"action": BaseModerationActions.STOP,
"threshold":0.5
},
"intent":{
"action": BaseModerationActions.STOP,
"threshold":0.5
}
}
comp_moderation_with_config = AmazonComprehendModerationChain(
moderation_config=moderation_config, #specify the configuration
client=comprehend_client, #optionally pass the Boto3 Client
verbose=True
)
template = """Question: {question}
Answer:"""
prompt = PromptTemplate(template=template, input_variables=["question"])
responses = [
"Final Answer: A credit card number looks like 1289-2321-1123-2387. A fake SSN number looks like 323-22-9980. John Doe's phone number is (999)253-9876.",
"Final Answer: This is a really shitty way of constructing a birdhouse. This is fucking insane to think that any birds would actually create their motherfucking nests here."
]
llm = FakeListLLM(responses=responses)
llm_chain = LLMChain(prompt=prompt, llm=llm)
chain = (
prompt
| comp_moderation_with_config
| {llm_chain.input_keys[0]: lambda x: x['output'] }
| llm_chain
| { "input": lambda x: x['text'] }
| comp_moderation_with_config
)
response = chain.invoke({"question": "A sample SSN number looks like this 123-456-7890. Can you give me some more samples?"})
print(response['output'])
```
### Output
```
> Entering new AmazonComprehendModerationChain chain...
Running AmazonComprehendModerationChain...
Running pii validation...
Found PII content..stopping..
The prompt contains PII entities and cannot be processed
```
---------
Co-authored-by: Piyush Jain <piyushjain@duck.com>
Co-authored-by: Anjan Biswas <anjanavb@amazon.com>
Co-authored-by: Jha <nikjha@amazon.com>
Co-authored-by: Bagatur <baskaryan@gmail.com>
This PR fixes `QuestionListOutputParser` text splitting.
`QuestionListOutputParser` incorrectly splits numbered list text into
lines. If text doesn't end with `\n` , the regex doesn't capture the
last item. So it always returns `n - 1` items, and
`WebResearchRetriever.llm_chain` generates less queries than requested
in the search prompt.
How to reproduce:
```python
from langchain.retrievers.web_research import QuestionListOutputParser
parser = QuestionListOutputParser()
good = parser.parse(
"""1. This is line one.
2. This is line two.
""" # <-- !
)
bad = parser.parse(
"""1. This is line one.
2. This is line two.""" # <-- No new line.
)
assert good.lines == ['1. This is line one.\n', '2. This is line two.\n'], good.lines
assert bad.lines == ['1. This is line one.\n', '2. This is line two.'], bad.lines
```
NOTE: Last item will not contain a line break but this seems ok because
the items are stripped in the
`WebResearchRetriever.clean_search_query()`.