This notebook showcases how to implement a multi-agent simulation where
a privileged agent decides who to speak.
This follows the polar opposite selection scheme as [multi-agent
decentralized speaker
selection](https://python.langchain.com/en/latest/use_cases/agent_simulations/multiagent_bidding.html).
We show an example of this approach in the context of a fictitious
simulation of a news network. This example will showcase how we can
implement agents that
- think before speaking
- terminate the conversation
This notebook showcases how to implement a multi-agent simulation
without a fixed schedule for who speaks when. Instead the agents decide
for themselves who speaks. We can implement this by having each agent
bid to speak. Whichever agent's bid is the highest gets to speak.
We will show how to do this in the example below that showcases a
fictitious presidential debate.
Tools for Bing, DDG and Google weren't consistent even though the
underlying implementations were.
All three services now have the same tools and implementations to easily
switch and experiment when building chains.
This notebook shows how the DialogueAgent and DialogueSimulator class
make it easy to extend the [Two-Player Dungeons & Dragons
example](https://python.langchain.com/en/latest/use_cases/agent_simulations/two_player_dnd.html)
to multiple players.
The main difference between simulating two players and multiple players
is in revising the schedule for when each agent speaks
To this end, we augment DialogueSimulator to take in a custom function
that determines the schedule of which agent speaks. In the example
below, each character speaks in round-robin fashion, with the
storyteller interleaved between each player.
I would like to contribute with a jupyter notebook example
implementation of an AI Sales Agent using `langchain`.
The bot understands the conversation stage (you can define your own
stages fitting your needs)
using two chains:
1. StageAnalyzerChain - takes context and LLM decides what part of sales
conversation is one in
2. SalesConversationChain - generate next message
Schema:
https://images-genai.s3.us-east-1.amazonaws.com/architecture2.png
my original repo: https://github.com/filip-michalsky/SalesGPT
This example creates a sales person named Ted Lasso who is trying to
sell you mattresses.
Happy to update based on your feedback.
Thanks, Filip
https://twitter.com/FilipMichalsky
Simplifies the [Two Agent
D&D](https://python.langchain.com/en/latest/use_cases/agent_simulations/two_player_dnd.html)
example with a cleaner, simpler interface that is extensible for
multiple agents.
`DialogueAgent`:
- `send()`: applies the chatmodel to the message history and returns the
message string
- `receive(name, message)`: adds the `message` spoken by `name` to
message history
The `DialogueSimulator` class takes a list of agents. At each step, it
performs the following:
1. Select the next speaker
2. Calls the next speaker to send a message
3. Broadcasts the message to all other agents
4. Update the step counter.
The selection of the next speaker can be implemented as any function,
but in this case we simply loop through the agents.
In this notebook, we show how we can use concepts from
[CAMEL](https://www.camel-ai.org/) to simulate a role-playing game with
a protagonist and a dungeon master. To simulate this game, we create a
`TwoAgentSimulator` class that coordinates the dialogue between the two
agents.
Improvements
* set default num_workers for ingestion to 0
* upgraded notebooks for avoiding dataset creation ambiguity
* added `force_delete_dataset_by_path`
* bumped deeplake to 3.3.0
* creds arg passing to deeplake object that would allow custom S3
Notes
* please double check if poetry is not messed up (thanks!)
Asks
* Would be great to create a shared slack channel for quick questions
---------
Co-authored-by: Davit Buniatyan <d@activeloop.ai>
The following calls were throwing an exception:
575b717d10/docs/use_cases/evaluation/agent_vectordb_sota_pg.ipynb (L192)575b717d10/docs/use_cases/evaluation/agent_vectordb_sota_pg.ipynb (L239)
Exception:
```
---------------------------------------------------------------------------
ValidationError Traceback (most recent call last)
Cell In[14], line 1
----> 1 chain_sota = RetrievalQA.from_chain_type(llm=OpenAI(temperature=0), chain_type="stuff", retriever=vectorstore_sota, input_key="question")
File ~/github/langchain/venv/lib/python3.9/site-packages/langchain/chains/retrieval_qa/base.py:89, in BaseRetrievalQA.from_chain_type(cls, llm, chain_type, chain_type_kwargs, **kwargs)
85 _chain_type_kwargs = chain_type_kwargs or {}
86 combine_documents_chain = load_qa_chain(
87 llm, chain_type=chain_type, **_chain_type_kwargs
88 )
---> 89 return cls(combine_documents_chain=combine_documents_chain, **kwargs)
File ~/github/langchain/venv/lib/python3.9/site-packages/pydantic/main.py:341, in pydantic.main.BaseModel.__init__()
ValidationError: 1 validation error for RetrievalQA
retriever
instance of BaseRetriever expected (type=type_error.arbitrary_type; expected_arbitrary_type=BaseRetriever)
```
The vectorstores had to be converted to retrievers:
`vectorstore_sota.as_retriever()` and `vectorstore_pg.as_retriever()`.
The PR also:
- adds the file `paul_graham_essay.txt` referenced by this notebook
- adds to gitignore *.pkl and *.bin files that are generated by this
notebook
Interestingly enough, the performance of the prediction greatly
increased (new version of langchain or ne version of OpenAI models since
the last run of the notebook): from 19/33 correct to 28/33 correct!
- Most important - fixes the relevance_fn name in the notebook to align
with the docs
- Updates comments for the summary:
<img width="787" alt="image"
src="https://user-images.githubusercontent.com/130414180/232520616-2a99e8c3-a821-40c2-a0d5-3f3ea196c9bb.png">
- The new conversation is a bit better, still unfortunate they try to
schedule a followup.
- Rm the max dialogue turns argument to the conversation function
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
Minor cosmetic changes
- Activeloop environment cred authentication in notebooks with
`getpass.getpass` (instead of CLI which not always works)
- much faster tests with Deep Lake pytest mode on
- Deep Lake kwargs pass
Notes
- I put pytest environment creds inside `vectorstores/conftest.py`, but
feel free to suggest a better location. For context, if I put in
`test_deeplake.py`, `ruff` doesn't let me to set them before import
deeplake
---------
Co-authored-by: Davit Buniatyan <d@activeloop.ai>
The doc loaders index was picking up a bunch of subheadings because I
mistakenly made the MD titles H1s. Fixed that.
also the easy minor warnings from docs_build
This PR proposes
- An NLAToolkit method to instantiate from an AI Plugin URL
- A notebook that shows how to use that alongside an example of using a
Retriever object to lookup specs and route queries to them on the fly
---------
Co-authored-by: Harrison Chase <hw.chase.17@gmail.com>
This PR adds a LangChain implementation of CAMEL role-playing example:
https://github.com/lightaime/camel.
I am sorry that I am not that familiar with LangChain. So I only
implement it in a naive way. There may be a better way to implement it.
Improvements to Deep Lake Vector Store
- much faster view loading of embeddings after filters with
`fetch_chunks=True`
- 2x faster ingestion
- use np.float32 for embeddings to save 2x storage, LZ4 compression for
text and metadata storage (saves up to 4x storage for text data)
- user defined functions as filters
Docs
- Added retriever full example for analyzing twitter the-algorithm
source code with GPT4
- Added a use case for code analysis (please let us know your thoughts
how we can improve it)
---------
Co-authored-by: Davit Buniatyan <d@activeloop.ai>
