- [ ] **PR title**: "experimental: Adding compatibility for
OllamaFunctions with ImagePromptTemplate"
- [ ] **PR message**:
- **Description:** Removes the outdated
`_convert_messages_to_ollama_messages` method override in the
`OllamaFunctions` class to ensure that ollama multimodal models can be
invoked with an image.
- **Issue:** #24174
---------
Co-authored-by: Joel Akeret <joel.akeret@ti&m.com>
Co-authored-by: Isaac Francisco <78627776+isahers1@users.noreply.github.com>
Co-authored-by: isaac hershenson <ihershenson@hmc.edu>
* Please see security warning already in existing class.
* The approach here is fundamentally insecure as it's relying on a block
approach rather than an approach based on only running allowed nodes.
So users should only use this code if its running from a properly
sandboxed environment.
- Description: Modified the prompt created by the function
`create_unstructured_prompt` (which is called for LLMs that do not
support function calling) by adding conditional checks that verify if
restrictions on entity types and rel_types should be added to the
prompt. If the user provides a sufficiently large text, the current
prompt **may** fail to produce results in some LLMs. I have first seen
this issue when I implemented a custom LLM class that did not support
Function Calling and used Gemini 1.5 Pro, but I was able to replicate
this issue using OpenAI models.
By loading a sufficiently large text
```python
from langchain_community.llms import Ollama
from langchain_openai import ChatOpenAI, OpenAI
from langchain_core.prompts import PromptTemplate
import re
from langchain_experimental.graph_transformers import LLMGraphTransformer
from langchain_core.documents import Document
with open("texto-longo.txt", "r") as file:
full_text = file.read()
partial_text = full_text[:4000]
documents = [Document(page_content=partial_text)] # cropped to fit GPT 3.5 context window
```
And using the chat class (that has function calling)
```python
chat_openai = ChatOpenAI(model="gpt-3.5-turbo", model_kwargs={"seed": 42})
chat_gpt35_transformer = LLMGraphTransformer(llm=chat_openai)
graph_from_chat_gpt35 = chat_gpt35_transformer.convert_to_graph_documents(documents)
```
It works:
```
>>> print(graph_from_chat_gpt35[0].nodes)
[Node(id="Jesu, Joy of Man's Desiring", type='Music'), Node(id='Godel', type='Person'), Node(id='Johann Sebastian Bach', type='Person'), Node(id='clever way of encoding the complicated expressions as numbers', type='Concept')]
```
But if you try to use the non-chat LLM class (that does not support
function calling)
```python
openai = OpenAI(
model="gpt-3.5-turbo-instruct",
max_tokens=1000,
)
gpt35_transformer = LLMGraphTransformer(llm=openai)
graph_from_gpt35 = gpt35_transformer.convert_to_graph_documents(documents)
```
It uses the prompt that has issues and sometimes does not produce any
result
```
>>> print(graph_from_gpt35[0].nodes)
[]
```
After implementing the changes, I was able to use both classes more
consistently:
```shell
>>> chat_gpt35_transformer = LLMGraphTransformer(llm=chat_openai)
>>> graph_from_chat_gpt35 = chat_gpt35_transformer.convert_to_graph_documents(documents)
>>> print(graph_from_chat_gpt35[0].nodes)
[Node(id="Jesu, Joy Of Man'S Desiring", type='Music'), Node(id='Johann Sebastian Bach', type='Person'), Node(id='Godel', type='Person')]
>>> gpt35_transformer = LLMGraphTransformer(llm=openai)
>>> graph_from_gpt35 = gpt35_transformer.convert_to_graph_documents(documents)
>>> print(graph_from_gpt35[0].nodes)
[Node(id='I', type='Pronoun'), Node(id="JESU, JOY OF MAN'S DESIRING", type='Song'), Node(id='larger memory', type='Memory'), Node(id='this nice tree structure', type='Structure'), Node(id='how you can do it all with the numbers', type='Process'), Node(id='JOHANN SEBASTIAN BACH', type='Composer'), Node(id='type of structure', type='Characteristic'), Node(id='that', type='Pronoun'), Node(id='we', type='Pronoun'), Node(id='worry', type='Verb')]
```
The results are a little inconsistent because the GPT 3.5 model may
produce incomplete json due to the token limit, but that could be solved
(or mitigated) by checking for a complete json when parsing it.
Pydantic allows empty strings:
```
from langchain.pydantic_v1 import Field, BaseModel
class Property(BaseModel):
"""A single property consisting of key and value"""
key: str = Field(..., description="key")
value: str = Field(..., description="value")
x = Property(key="", value="")
```
Which can produce errors downstream. We simply ignore those records
```SemanticChunker``` currently provide three methods to split the texts semantically:
- percentile
- standard_deviation
- interquartile
I propose new method ```gradient```. In this method, the gradient of distance is used to split chunks along with the percentile method (technically) . This method is useful when chunks are highly correlated with each other or specific to a domain e.g. legal or medical. The idea is to apply anomaly detection on gradient array so that the distribution become wider and easy to identify boundaries in highly semantic data.
I have tested this merge on a set of 10 domain specific documents (mostly legal).
Details :
- **Issue:** Improvement
- **Dependencies:** NA
- **Twitter handle:** [x.com/prajapat_ravi](https://x.com/prajapat_ravi)
@hwchase17
---------
Co-authored-by: Raviraj Prajapat <raviraj.prajapat@sirionlabs.com>
Co-authored-by: isaac hershenson <ihershenson@hmc.edu>
- **Description:**
The generated relationships in the graph had no properties, but the
Relationship class was properly defined with properties. This made it
very difficult to transform conditional sentences into a graph. Adding
properties to relationships can solve this issue elegantly.
The changes expand on the existing LLMGraphTransformer implementation
but add the possibility to define allowed relationship properties like
this: LLMGraphTransformer(llm=llm, relationship_properties=["Condition",
"Time"],)
- **Issue:**
no issue found
- **Dependencies:**
n/a
- **Twitter handle:**
@IstvanSpace
-Quick Test
=================================================================
from dotenv import load_dotenv
import os
from langchain_community.graphs import Neo4jGraph
from langchain_experimental.graph_transformers import
LLMGraphTransformer
from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.documents import Document
load_dotenv()
os.environ["NEO4J_URI"] = os.getenv("NEO4J_URI")
os.environ["NEO4J_USERNAME"] = os.getenv("NEO4J_USERNAME")
os.environ["NEO4J_PASSWORD"] = os.getenv("NEO4J_PASSWORD")
graph = Neo4jGraph()
llm = ChatOpenAI(temperature=0, model_name="gpt-4o")
llm_transformer = LLMGraphTransformer(llm=llm)
#text = "Harry potter likes pies, but only if it rains outside"
text = "Jack has a dog named Max. Jack only walks Max if it is sunny
outside."
documents = [Document(page_content=text)]
llm_transformer_props = LLMGraphTransformer(
llm=llm,
relationship_properties=["Condition"],
)
graph_documents_props =
llm_transformer_props.convert_to_graph_documents(documents)
print(f"Nodes:{graph_documents_props[0].nodes}")
print(f"Relationships:{graph_documents_props[0].relationships}")
graph.add_graph_documents(graph_documents_props)
---------
Co-authored-by: Istvan Lorincz <istvan.lorincz@pm.me>
Co-authored-by: Bagatur <22008038+baskaryan@users.noreply.github.com>
Remove the REPL from community, and suggest an alternative import from
langchain_experimental.
Fix for this issue:
https://github.com/langchain-ai/langchain/issues/14345
This is not a bug in the code or an actual security risk. The python
REPL itself is behaving as expected.
The PR is done to appease blanket security policies that are just
looking for the presence of exec in the code.
---------
Co-authored-by: Erick Friis <erick@langchain.dev>
This PR addresses several lint errors in the core package of LangChain.
Specifically, the following issues were fixed:
1.Unexpected keyword argument "required" for "Field" [call-arg]
2.tests/integration_tests/chains/test_cpal.py:263: error: Unexpected
keyword argument "narrative_input" for "QueryModel" [call-arg]
This should make it obvious that a few of the agents in langchain
experimental rely on the python REPL as a tool under the hood, and will
force users to opt-in.
We need to use a different version of numpy for py3.8 and py3.12 in
pyproject.
And so do projects that use that Python version range and import
langchain.
- **Twitter handle:** _cbornet
They cause `poetry lock` to take a ton of time, and `uv pip install` can
resolve the constraints from these toml files in trivial time
(addressing problem with #19153)
This allows us to properly upgrade lockfile dependencies moving forward,
which revealed some issues that were either fixed or type-ignored (see
file comments)
0.2 is not a breaking release for core (but it is for langchain and
community)
To keep the core+langchain+community packages in sync at 0.2, we will
relax deps throughout the ecosystem to tolerate `langchain-core` 0.2
0.2rc
migrations
- [x] Move memory
- [x] Move remaining retrievers
- [x] graph_qa chains
- [x] some dependency from evaluation code potentially on math utils
- [x] Move openapi chain from `langchain.chains.api.openapi` to
`langchain_community.chains.openapi`
- [x] Migrate `langchain.chains.ernie_functions` to
`langchain_community.chains.ernie_functions`
- [x] migrate `langchain/chains/llm_requests.py` to
`langchain_community.chains.llm_requests`
- [x] Moving `langchain_community.cross_enoders.base:BaseCrossEncoder`
->
`langchain_community.retrievers.document_compressors.cross_encoder:BaseCrossEncoder`
(namespace not ideal, but it needs to be moved to `langchain` to avoid
circular deps)
- [x] unit tests langchain -- add pytest.mark.community to some unit
tests that will stay in langchain
- [x] unit tests community -- move unit tests that depend on community
to community
- [x] mv integration tests that depend on community to community
- [x] mypy checks
Other todo
- [x] Make deprecation warnings not noisy (need to use warn deprecated
and check that things are implemented properly)
- [x] Update deprecation messages with timeline for code removal (likely
we actually won't be removing things until 0.4 release) -- will give
people more time to transition their code.
- [ ] Add information to deprecation warning to show users how to
migrate their code base using langchain-cli
- [ ] Remove any unnecessary requirements in langchain (e.g., is
SQLALchemy required?)
---------
Co-authored-by: Erick Friis <erick@langchain.dev>
Before you could only extract triples (diffbot calls it facts) from
diffbot to avoid isolated nodes. However, sometimes isolated nodes can
still be useful like for prefiltering, so we want to allow users to
extract them if they want. Default behaviour is unchanged.
Update python.py（experimental：Added code for PythonREPL）
Added code for PythonREPL, defining a static method 'sanitize_input'
that takes the string 'query' as input and returns a sanitizing string.
The purpose of this method is to remove unwanted characters from the
input string, Specifically:
1. Delete the whitespace at the beginning and end of the string (' \s').
2. Remove the quotation marks (`` ` ``) at the beginning and end of the
string.
3. Remove the keyword "python" at the beginning of the string (case
insensitive) because the user may have typed it.
This method uses regular expressions (regex) to implement sanitizing.
It all started with this code：
from langchain.agents import Tool
from langchain_experimental.utilities import PythonREPL
python_repl = PythonREPL()
repl_tool = Tool(
name="python_repl",
description="Remove redundant formatting marks at the beginning and end
of source code from input.Use a Python shell to execute python commands.
If you want to see the output of a value, you should print it out with
`print(...)`.",
func=python_repl.run,
)
When I call the agent to write a piece of code for me and execute it
with the defined code, I must get an error: SyntaxError('invalid
syntax', ('<string>', 1, 1,'In', 1, 2))
After checking, I found that pythonREPL has less formatting of input
code than the soon-to-be deprecated pythonREPL tool, so I added this
step to it, so that no matter what code I ask the agent to write for me,
it can be executed smoothly and get the output result.
I have tried modifying the prompt words to solve this problem before,
but it did not work, and by adding a simple format check, the problem is
well resolved.
<img width="1271" alt="image"
src="https://github.com/langchain-ai/langchain/assets/164149097/c49a685f-d246-4b11-b655-fd952fc2f04c">
---------
Co-authored-by: Bagatur <22008038+baskaryan@users.noreply.github.com>
Co-authored-by: Bagatur <baskaryan@gmail.com>
Implemented bind_tools for OllamaFunctions.
Made OllamaFunctions sub class of ChatOllama.
Implemented with_structured_output for OllamaFunctions.
integration unit test has been updated.
notebook has been updated.
---------
Co-authored-by: Bagatur <baskaryan@gmail.com>
**Description:** Fixes a bug in the HuggingGPT task execution logic
here:
except Exception as e:
self.status = "failed"
self.message = str(e)
self.status = "completed"
self.save_product()
where a caught exception effectively just sets `self.message` and can
then throw an exception if, e.g., `self.product` is not defined.
**Issue:** None that I'm aware of.
**Dependencies:** None
**Twitter handle:** https://twitter.com/michaeljschock
Co-authored-by: Bagatur <22008038+baskaryan@users.noreply.github.com>
- **Description:** Currently, the regex is static (`r"(?<=[.?!])\s+"`),
which is only useful for certain use cases. The current change only
moves this to be a parameter of split_text(). Which adds flexibility
without making it more complex (as the default regex is still the same).
- **Issue:** Not applicable (I searched, no one seems to have created
this issue yet).
- **Dependencies:** None.
_If no one reviews your PR within a few days, please @-mention one of
baskaryan, efriis, eyurtsev, hwchase17._
---------
Co-authored-by: Bagatur <22008038+baskaryan@users.noreply.github.com>
Co-authored-by: Bagatur <baskaryan@gmail.com>
