### Background
Continuing to implement all the interface methods defined by the
`VectorStore` class. This PR pertains to implementation of the
`max_marginal_relevance_search` method.
### Changes
- a `max_marginal_relevance_search` method implementation has been added
in `weaviate.py`
- tests have been added to the the new method
- vcr cassettes have been added for the weaviate tests
### Test Plan
Added tests for the `max_marginal_relevance_search` implementation
### Change Safety
- [x] I have added tests to cover my changes
- Modify SVMRetriever class to add an optional relevancy_threshold
- Modify SVMRetriever.get_relevant_documents method to filter out
documents with similarity scores below the relevancy threshold
- Normalized the similarities to be between 0 and 1 so the
relevancy_threshold makes more sense
- The number of results are limited to the top k documents or the
maximum number of relevant documents above the threshold, whichever is
smaller
This code will now return the top self.k results (or less, if there are
not enough results that meet the self.relevancy_threshold criteria).
The svm.LinearSVC implementation in scikit-learn is non-deterministic,
which means
SVMRetriever.from_texts(["bar", "world", "foo", "hello", "foo bar"])
could return [3 0 5 4 2 1] instead of [0 3 5 4 2 1] with a query of
"foo".
If you pass in multiple "foo" texts, the order could be different each
time. Here, we only care if the 0 is the first element, otherwise it
will offset the text and similarities.
Example:
```python
retriever = SVMRetriever.from_texts(
["foo", "bar", "world", "hello", "foo bar"],
OpenAIEmbeddings(),
k=4,
relevancy_threshold=.25
)
result = retriever.get_relevant_documents("foo")
```
yields
```python
[Document(page_content='foo', metadata={}), Document(page_content='foo bar', metadata={})]
```
---------
Co-authored-by: Brandon Sandoval <52767641+account00001@users.noreply.github.com>
re
https://github.com/hwchase17/langchain/issues/439#issuecomment-1510442791
I think it's not polite for a library to use the root logger
both of these forms are also used:
```
logger = logging.getLogger(__name__)
logger = logging.getLogger(__file__)
```
I am not sure if there is any reason behind one vs the other? (...I am
guessing maybe just contributed by different people)
it seems to me it'd be better to consistently use
`logging.getLogger(__name__)`
this makes it easier for consumers of the library to set up log
handlers, e.g. for everything with `langchain.` prefix
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
Last week I added the `PDFMinerPDFasHTMLLoader`. I am adding some
example code in the notebook to serve as a tutorial for how that loader
can be used to create snippets of a pdf that are structured within
sections. All the other loaders only provide the `Document` objects
segmented by pages but that's pretty loose given the amount of other
metadata that can be extracted.
With the new loader, one can leverage font-size of the text to decide
when a new sections starts and can segment the text more semantically as
shown in the tutorial notebook. The cell shows that we are able to find
the content of entire section under **Related Work** for the example pdf
which is spread across 2 pages and hence is stored as two separate
documents by other loaders
Fixes a bug I was seeing when the `TokenTextSplitter` was correctly
splitting text under the gpt3.5-turbo token limit, but when firing the
prompt off too openai, it'd come back with an error that we were over
the context limit.
gpt3.5-turbo and gpt-4 use `cl100k_base` tokenizer, and so the counts
are just always off with the default `gpt-2` encoder.
It's possible to pass along the encoding to the `TokenTextSplitter`, but
it's much simpler to pass the model name of the LLM. No more concern
about keeping the tokenizer and llm model in sync :)
I got the following stacktrace when the agent was trying to search
Wikipedia with a huge query:
```
Thought:{
"action": "Wikipedia",
"action_input": "Outstanding is a song originally performed by the Gap Band and written by member Raymond Calhoun. The song originally appeared on the group's platinum-selling 1982 album Gap Band IV. It is one of their signature songs and biggest hits, reaching the number one spot on the U.S. R&B Singles Chart in February 1983. \"Outstanding\" peaked at number 51 on the Billboard Hot 100."
}
Traceback (most recent call last):
File "/usr/src/app/tests/chat.py", line 121, in <module>
answer = agent_chain.run(input=question)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/site-packages/langchain/chains/base.py", line 216, in run
return self(kwargs)[self.output_keys[0]]
^^^^^^^^^^^^
File "/usr/local/lib/python3.11/site-packages/langchain/chains/base.py", line 116, in __call__
raise e
File "/usr/local/lib/python3.11/site-packages/langchain/chains/base.py", line 113, in __call__
outputs = self._call(inputs)
^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/site-packages/langchain/agents/agent.py", line 828, in _call
next_step_output = self._take_next_step(
^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/site-packages/langchain/agents/agent.py", line 725, in _take_next_step
observation = tool.run(
^^^^^^^^^
File "/usr/local/lib/python3.11/site-packages/langchain/tools/base.py", line 73, in run
raise e
File "/usr/local/lib/python3.11/site-packages/langchain/tools/base.py", line 70, in run
observation = self._run(tool_input)
^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/site-packages/langchain/agents/tools.py", line 17, in _run
return self.func(tool_input)
^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/site-packages/langchain/utilities/wikipedia.py", line 40, in run
search_results = self.wiki_client.search(query)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/site-packages/wikipedia/util.py", line 28, in __call__
ret = self._cache[key] = self.fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.11/site-packages/wikipedia/wikipedia.py", line 109, in search
raise WikipediaException(raw_results['error']['info'])
wikipedia.exceptions.WikipediaException: An unknown error occured: "Search request is longer than the maximum allowed length. (Actual: 373; allowed: 300)". Please report it on GitHub!
```
This commit limits the maximum size of the query passed to Wikipedia to
avoid this issue.
This allows to adjust the number of results to retrieve and filter
documents based on metadata.
---------
Co-authored-by: Altay Sansal <altay.sansal@tgs.com>
Add a method that exposes a similarity search with corresponding
normalized similarity scores. Implement only for FAISS now.
### Motivation:
Some memory definitions combine `relevance` with other scores, like
recency , importance, etc.
While many (but not all) of the `VectorStore`'s expose a
`similarity_search_with_score` method, they don't all interpret the
units of that score (depends on the distance metric and whether or not
the the embeddings are normalized).
This PR proposes a `similarity_search_with_normalized_similarities`
method that lets consumers of the vector store not have to worry about
the metric and embedding scale.
*Most providers default to euclidean distance, with Pinecone being one
exception (defaults to cosine _similarity_).*
---------
Co-authored-by: Harrison Chase <hw.chase.17@gmail.com>
The encoding fetch was out of date. Luckily OpenAI has a nice[
`encoding_for_model`](46287bfa49/tiktoken/model.py)
function in `tiktoken` we can use now.
Title, lang and description are on almost every web page, and are
incredibly useful pieces of information that currently isn't captured
with the current web base loader
I thought about adding the title and description to the content of the
document, as
that content could be useful in search, but I left it out for right now.
If you think
it'd be worth adding, happy to add it.
I've found it's nice to have the title/description in the metadata to
have some structured data
when retrieving rows from vectordbs for use with summary and source
citation, so if we do want to add it to the `page_content`, i'd advocate
for it to also be included in metadata.
Add SVM retriever class, based on
https://github.com/karpathy/randomfun/blob/master/knn_vs_svm.ipynb.
Testing still WIP, but the logic is correct (I have a local
implementation outside of Langchain working).
---------
Co-authored-by: Lance Martin <122662504+PineappleExpress808@users.noreply.github.com>
Co-authored-by: rlm <31treehaus@31s-MacBook-Pro.local>
Same as similarity_search, allows child classes to add vector
store-specific args (this was technically already happening in couple
places but now typing is correct).
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>
Note to self: Always run integration tests, even on "that last minute
change you thought would be safe" :)
---------
Co-authored-by: Mike Lambert <mike.lambert@anthropic.com>
**About**
Specify encoding to avoid UnicodeDecodeError when reading .txt for users
who are following the tutorial.
**Reference**
```
return codecs.charmap_decode(input,self.errors,decoding_table)[0]
UnicodeDecodeError: 'charmap' codec can't decode byte 0x9d in position 1205: character maps to <undefined>
```
**Environment**
OS: Win 11
Python: 3.8
* Adds an Anthropic ChatModel
* Factors out common code in our LLMModel and ChatModel
* Supports streaming llm-tokens to the callbacks on a delta basis (until
a future V2 API does that for us)
* Some fixes
Allows users to specify what files should be loaded instead of
indiscriminately loading the entire repo.
extends #2851
NOTE: for reviewers, `hide whitespace` option recommended since I
changed the indentation of an if-block to use `continue` instead so it
looks less like a Christmas tree :)
Mentioned the idea here initially:
https://github.com/hwchase17/langchain/pull/2106#issuecomment-1487509106
Since there have been dialect-specific issues, we should use
dialect-specific prompts. This way, each prompt can be separately
modified to best suit each dialect as needed. This adds a prompt for
each dialect supported in sqlalchemy (mssql, mysql, mariadb, postgres,
oracle, sqlite). For this initial implementation, the only differencse
between the prompts is the instruction for the clause to use to limit
the number of rows queried for, and the instruction for wrapping column
names using each dialect's identifier quote character.
Optimization :Limit search results when k < 10
Fix issue when k > 10: Elasticsearch will return only 10 docs
[default-search-result](https://www.elastic.co/guide/en/elasticsearch/reference/current/paginate-search-results.html)
By default, searches return the top 10 matching hits
Add size parameter to the search request to limit the number of returned
results from Elasticsearch. Remove slicing of the hits list, since the
response will already contain the desired number of results.
