* Extract optional dependencies
* Separate local mode into llms-llama-cpp and embeddings-huggingface for clarity
* Support Ollama embeddings
* Upgrade to llamaindex 0.10.14. Remove legacy use of ServiceContext in ContextChatEngine
* Fix vector retriever filters
Updated ui.py to include a small sleep timer while building the stream deltas. This recursive function fires off so quickly to eats up too much of the CPU. This small sleep frees up the CPU to not be bottlenecked. This value can go lower/shorter. But 0.02 or 0.025 seems to work well. (#1589)
Co-authored-by: root <root@wesgitlabdemo.icl.gtri.org>
This mode behaves the same as the openai mode, except that it allows setting custom models not
supported by OpenAI. It can be used with any tool that serves models from an OpenAI compatible API.
Implements #1424
As discussed on Discord, the decision has been made to remove the system prompts by default, to better segregate the API and the UI usages.
A concurrent PR (#1353) is enabling the dynamic setting of a system prompt in the UI.
Therefore, if UI users want to use a custom system prompt, they can specify one directly in the UI.
If the API users want to use a custom prompt, they can pass it directly into their messages that they are passing to the API.
In the highlight of the two use case above, it becomes clear that default system_prompt does not need to exist.
* Fix the parallel ingestion mode, and make it available through conf
Also updated the documentation to show how to configure the ingest mode.
* PR feedback: redirect to documentation
* added max_new_tokens as a configuration option to the llm block in settings
* Update fern/docs/pages/manual/settings.mdx
Co-authored-by: lopagela <lpglm@orange.fr>
* Update private_gpt/settings/settings.py
Add default value for max_new_tokens = 256
Co-authored-by: lopagela <lpglm@orange.fr>
* Addressed location of docs comment
* reformatting from running 'make check'
* remove default config value from settings.yaml
---------
Co-authored-by: lopagela <lpglm@orange.fr>
* Update llama_index to 0.9.3
Had to change some imports because of breaking change durin the llama_index update to 0.9.0
* Update poetry.lock after update of llama_index
* Add simple Basic auth
To enable the basic authentication, one must set `server.auth.enabled`
to true.
The static string defined in `server.auth.secret` must be set in the
header `Authorization`.
The health check endpoint will always be accessible, no matter the API
auth configuration.
* Fix linting and type check
* Fighting with mypy being too restrictive
Had to disable mypy in the `auth` as we are not using the same signature
for the authenticated method.
mypy was complaining that the signatures of `authenticated` must be
identical, no matter in which logical branch we are.
Given that fastapi is accomodating itself of method signatures (it will
inject the dependencies in the method call), this warning of mypy is
actually preventing us to do something legit.
mypy doc: https://mypy.readthedocs.io/en/stable/common_issues.html
* Write tests to verify that the simple auth is working
* Added logs at generation of the UI, and generate the UI in an object
* Make ingest script more verbose in case of an error at ingestion time
* Removed the explicit state in the UI containing ingested files
* Make script of ingestion a bit more verbose by displaying stack traces
* Change the browser tab title of privateGPT ui to `My Private GPT`
Patch the default list of llama_index to support JSON files.
This injection of JSON documents should improve the comprehension in
JSON files, as there is a parsing of JSON files.
A file that is ingested will be transformed into several documents (that
are organized into nodes).
This endpoint is deleting documents (bits of a file). These bits can be
retrieved thanks to the endpoint to list all the documents.
* Disable Gradio Analytics
Gradio analytics can be disabled by either using the kwargs `enable_analytics` on `gr.Blocks`, or by setting the env variable `GRADIO_ANALYTICS_ENABLED` to something different from `True`.
Since that Gradio does not seem to respect their code contract (around `enable_analytics`), and that they are performing other operations only based on the value of `GRADIO_ANALYTICS_ENABLED` (c.f. `gradio.strings` https://github.com/gradio-app/gradio/blob/main/gradio/strings.py#L39), we are disabling gradio analytics by setting the required env variable to `False`.
Note: Setting an environment variables using `os.environ['foo'] = 'bar'` on system that are not based on unix might not work.
c.f. https://docs.python.org/3/library/os.html#os.environ for details on how `os.environ` works and all its caveats
* Update private_gpt/__init__.py
* Configure simple builtin logging
Changed the 2 existing `print` in the `private_gpt` code base into actual python logging, stop using loguru (dependency will be dropped in a later commit).
Try to use the `key=value` logging convention in logs (to indicate what dynamic values represents, and what is dynamic vs not).
Using `%s` log style, so that the string formatting is pushed inside the logger, giving the ability to the logger to determine if the string need to be formatted or not (i.e. strings from debug logs might not be formatted if the log level is not debug)
The (basic) builtin log configuration have been placed in `private_gpt/__init__.py` in order to initialize the logging system even before we start to launch any python code in `private_gpt` package (ensuring we get any initialization log formatted as we want to)
Disabled `uvicorn` custom logging format, resulting in having uvicorn logs being outputted in our formatted.
Some more concise format could be used if we want to, especially:
```
COMPACT_LOG_FORMAT = '%(asctime)s.%(msecs)03d [%(levelname)s] %(name)s - %(message)s'
```
Python documentation and cookbook on logging for reference:
* https://docs.python.org/3/library/logging.html
* https://docs.python.org/3/howto/logging.html
* Removing loguru from the dependencies
Result of `poetry remove loguru`
* PR feedback: using `logger` variable name instead of `log`
---------
Co-authored-by: Louis Melchior <louis@jaris.io>
Iván Martínez
TQ
icsy7867
Ygal Blum
Naveen Kannan
CognitiveTech
Robert Gay
Matthew Hill
Rohit Das
3ly-13
Louis Melchior
Gianni Acquisto
Francisco García Sierra
Pablo Orgaz
Anush
Pierre Marais