Dandy now supports agentic processing through the LlmAgent module located in dandy.llm.
Works very similarly to a LlmBot but requires processors attribute that is a sequence of BaseProcessor subclasses.
All processors must have a description attribute to tell the agent how to use them and when.
Fixes
Fixed a bug on LlmBot that prevented you from using the intel_object argument in the process method.
BaseProcessor method process now maintains it's signature properly.
v0.18.0
Breaking
BaseListIntel now works by creating a single attribute in the subclass that must be typed as a list.
Features
The BaseLlmMap class now has a new required attribute map_keys_description that drastically improves accuracy.
Changes
On the BaseLlmMap.process the argument choice_count was renamed to max_return_values to make it more clear what it does and is now optional.
The BaseListIntel now requires you to set the type of the items class variable with a type hint.
Fixes
Changed Prompt into a dataclass to allow typing with BaseIntel, BaseListIntel and BaseModel.
Fixed a problem with the way BaseLlmMap was prompting for single responses using max_return_values=1.
LlmService has been refactored to handle retrying and progressive exceptions much better.
v0.17.0
Features
Added a new calculator module to dandy to help solve math problems related to using Dandy.
In dandy.calculator.llm_calculator you can find model_size_for_inference_to_vram_gb and model_size_and_token_count_for_inference_to_vram_gb
There is a quick shortcut in the dandy cli to model_size_and_token_count_for_inference_to_vram_gb using the -c flag.
v0.16.0
Features
Added a new MessageHistory class that can be used to submit a history of messages to be used in the request prompt.
Fixes
Corrected the RequestMessage to have a Literal['user', 'assistant', 'system'] type for the role field.
Fixed the formatting internally for messages inside the LlmService class.
v0.15.0
Features
The Recorder module now properly tracks token usage and run times.
New delete_all_recordings and delete_recording methods added to the Recorder class for memory management.
New get_recording method added to the Recorder class for easy access.
Changes
Move around and update elements in the html rendered output for the Recorder class.
Fixes
Updated Recording class to properly track token usage and run times.
Removed the broken assistant_str_prompt_to_str method from BaseLlmConfig and LlmService classes.
v0.14.6
Fixes
Rename EventItem to EventAttribute for more seperation of concerns when it comes to dictionaries.
v0.14.5
Fixes
Fixed a problem with the Recorder class not outputting strings.
v0.14.4
Breaking
All of the recorder decorators have been renamed to recorder_to_html_file, recorder_to_markdown_file and recorder_to_json_file.
Fixes
Remove relative path system from dandy as it caused to many edge case problems.
v0.14.3
Fixes
Fixed a problem with SqliteCache not handling cleared tables correctly.
Added the clear_all method to MemoryCache and SqliteCache to fix clearing caches properly.
Fixed recorder shortcut methods for rendering to files and strings.
v0.14.2
Fixes
Fixed the file method on the Prompt to properly hand relative paths.
The relative_parents argument makes it quick to traverse parent directories by count.
v0.14.1
Features
Added object_source method to Prompt to get the source of a specific object in a module.
Drastically improved the LLM service to retry and handle prompt correction.
Fixes
Fixed the dandy.utils to correct a circular import problem.
Fixed a problem with the file method on the Prompt class to search the settings.BASE_PATH directory.
v0.14.0
Breaking
The debug module is now the recorder module.
This also includes many refactors like DebugRecorder to Recorder ETC.
In settings ALLOW_DEBUG_RECORDING has been removed.
In settings ALLOW_RECORDING_TO_FILE hase been added with a default value of False.
The destroy method on Cache object has now been replaced with destroy_all for clearing all caches.
Features
DebugRecorder is now Recorder and has new rendering methods and overall improved rendering quality.
You can now render to json, html and markdown files and strings for easy exporting.
Changes
SqliteCache now uses one file to store all different caches.
this allows for more fine control in the same way the MemoryCache works.
Fixes
Fixed the llm service to properly handle retrying when an invalid response is provided with progressive exceptions.
Improved the BaseIntel class to have much better handling for complex operations.
v0.13.3
Breaking
Moved 'dandy.core.cache' to 'dandy.cache'
Fixes
Debug recorder correctly renders OpenAI service request messages.
Added more testing for more complex JSON Schemas.
Fixed a bug where the Prompt.prompt method would not accept strings properly.
v0.13.2
Fixes
Fixed problem with BaseIntel required fields using nested includes and excludes.
v0.13.1
Fixes
Fixed bug with using include and exclude on BaseIntel objects that did not validate filled and empty fields properly.
v0.13.0
Breaking
Removed the assistant_str_prompt_to_str method and sub methods from LlmService.
Everything should be processed through the LlmBot going forward.
Features
Added a new processor called LlmMap that uses the Map object to easily create contextual relationships to python objects.
Debug recorder now shows the JSON schema that was sent to the llm service.
New decorator called debug_recorder_to_html(debug_name: str) that allows you to easily wrap a function or method.
Changes
Removed the Contrib Selection llm bot.
Fixes
Fixed a caching to be more robust and understand changes with in dandy processors for more accurate caching.
Drastically improved the testing of the OpenAI llm service.
v0.12.0
Features
The BaseLlmBot now supports images and image_files in the process_prompt_to_intel method.
Make sure you have the config using a compatible llm model that supports vision.
Fixes
Fixed the hash key generation process to work with more data types and maintain better consistency.
Improved testing of the OpenAI llm service.
v0.11.3
Changes
Refactored internal project structure to get ready for next AI features (TTS, STT, VISION ETC).
Fixes
Fixed a bug with clearing non-existent or empty caches.
v0.11.2
Notes
This release was combined with v0.11.1 for a single release.
Features
Updated example with better use of Prompt objects.
Added to_markdown_file method for the Book class in the example.
Updated caching objects to be easier to clear.
dandy.cache.MemoryCache and dandy.cache.SqliteCache have class method clear and destroy.
Fixes
Added text to global service prompt to improve response quality.
Fix bug with updating non-flat intel objects.
v0.11.0
Breaking
All uses of the process_prompt_to_intel method now require you to specify either an intel_class or an intel_object argument.
Features
A new example has been created that is much easier to follow and showcases the features of Dandy.
Added a new Intel class called BaseListIntel that is used to create an iterable intel object that behaves like a list.
When using process_prompt_to_intel you can now submit a intel_class or intel_object.
Submitting a class will return you a new instance of the class.
Submitting the object will return you a modified copy of the object.
The method process_prompt_to_intel now supports include_fields and exclude_fields which allow you to only include or exclude fields from the intel object or class.
Caching is now supported through the cache_to_memory and cache_to_sqlite decorators.
Check out the dandy/default_settings.py file to see how to configure caching beyond the defaults.
Decorator argument cache_name which can be used to separate the cache objects / files, default is dandy.
Decorator argument limit which can be used to set an upper limit on the number of items that can be cached, default is in settings.
Changes
Removed the old examples (Cookie Recipe and Pirate Story)
Exceptions are now being divided into two categories: DandyCriticalException and DandyRecoverableException.
Both of this will inherit from DandyException.
The DandyRecoverableException will be used to allow developers to attempt recovering from exceptions safely.
The DandyCriticalException will be for when an exception is unrecoverable and must be handled.
Fixes
Update the process_to_intel method used throughout the project to properly reflect the postfix_system_prompt argument.
Added missing return to the __str__ method in the Url class (Thanks Pyright).
v0.10.0
Breaking
Renamed Bot to BaseBot
Renamed Workflow to BaseWorkflow
The LLM API for Ollama now only works with 0.5.0 or greater.
The LLM API for OpenAI now only works with gpt-4o-mini or greater.
Documentation
We have an initial working documentation website that can be viewed at https://dandysoftware.com
Our new website is powered by mkdocs, mkdocstrings, mkdocs-include-markdown-plugin and mkdocs-material.
Features
In the LLM_CONFIGS in your settings the TYPE, HOST, PORT AND API_KEY from the DEFAULT config will now flow to the other configs if they are not specificed.
Added --version to the CLI interface for Dandy.
The OpenAI llm service now use json_schema for the response format.
The OllamaAI llm service now use json_schema for the response format.
Changes
Rebuilt the document structure.
Fixes
Dandy CLI now properly create default settings based on your environment variables.
Fixed the way the settings are handled so they properly show error messages for missing settings.
v0.9.2
Fixes
Update requirements.txt
v0.9.1
Fixes
Added default instructions prompt to the LlmBot class.
v0.9.0
Features
LlmBots now have a default built in process method that takes in a prompt and returns an intel object.
Changed our http handling library to httpx.
The contrib choice llm bot has been replaced with the much simpler selector llm bot.
The Prompt class init now has a text argument that will create a prompt with a text snippet automatically for simple prompts.
New setting DEFAULT_LLM_REQUEST_TIME_OUT that controls the timeout for LLM requests default is None.
Changes
Moved llm_bot from dandy.bot to dandy.llm.bot to match our refactoring changes.
Changed the base class from Handler to BaseProcessor
Refactored Intel to BaseIntel to improve readability.
Added BaseLlmBot class to dandy.llm.bot to be used for creating all llm bots.
BaseLlmBot config now takes just a string that is one of the LLM_CONFIGS in the settings.
Fixes
There is now a DefaultLlmIntel class that is used as the default intel class for LlmBots that has one attribute called text.
Fixed a bunch of Generic Type handling through-out the project.
Connection retry count of zero no longer causes an error.
Refactor llm internal packages to match their usage better.
Fixed AsyncFuture to allow you to access the result after accessing it once.
Fixed CLI to properly load environment variables and settings in the correct order.
v0.8.1
Fixes
Fixed the settings module validation to be much easier to implement and debug.
Fixed but that accidentally overwrote settings files for the user.
v0.8.0
Major Changes
We have created a new class called Intel that is the pydantic BaseModel class renamed to give more separation of concerns between Dandy code and your code.
For the most part of this project the word Model has been refactored to Intel to create more separation of concerns in projects.
The word Model has a lot of meaning in the context of Dandy, artificial intelligence, databases, libraries and other frameworks.
Our hope is this creates a very clear line between these specific objects and the rest of your code.
v0.7.0
Major Improvement
All the changes in v0.7.0 should reduce the over all code required to work with Dandy by up to 50%.
Features
Project structure improvement with new settings file.
All projects going forward will require a dandy_settings.py file in the root of your project.
This file will also need a BASE_PATH str variable set to the root of your project.
This file will also need a LLM_CONFIGS dict variable with a DEFAULT llm config.
Debug recorder can now output to a json string or file.
Added randomize seed to LLM config that will randomize the seed every time the config is used.
Added new evaluate cli command for evaluating your Dandy code -e --evaluate.
Currently only supports Prompt evaluation.
ALLOW_DEBUG_RECORDING was added to the settings for project wide control of the debug recorder.
defaulted to False.
You can now select your llm config from the cli using the -l --llm-config flag.
Changes
Updated the readme to match the new project structure.
All settings and llm configs are now managed through the Dandy settings module.
The environment variable DANDY_SETTINGS_MODULE can be used to specify the settings module to be used.
The system will default to look for a dandy_settings.py file in the current working directory or sys.path.
Moved a lot of project wide constants into the const.py file.
Fixes
Fixed the user dandy_settings.py not loading properly with in the internal Dandy modules.
Fixed readme to match new project structure and configuration setup.
Fixed the cli to properly use the dandy_settings.py file in the current working directory.
Improved testing coverage across the whole project.
v0.6.1
Features
Added heading and sub headings to the prompt class.
Changes
Improved testing across the whole project.
Prompt assembly is more reliable and consistent if you want empty lines you need to use the .line_break method.
Fixes
Fixed lots of prompt formatting issues.
Futures properly raise an exception when they time out.
v0.6.0
Features
You can now cancel futures using the cancel method.
LLmBots can now be configured with max_input_tokens and max_output_tokens
Changes
Removed the role attribute from the llm bot class.
Updated the readme to match the new project structure.
Example changed to use more features and provide a better example of usage.
Refactored a lot of naming to be more clear and direct.
Fixes
Fixed futures to properly use the thread pool executor to wait for results.
Fixed the examples to properly match the new project structure.
Fixed bug with futures not timing out properly.
v0.5.2
Fixes
Fixed the bearer token usage for use with ollama and reverse proxy servers.
v0.5.0
Features
Improved the debug recorder html output to be more user friendly and easier to read.
Added ID's to all the tracked events in the debug recorder.
Changes
Changed the service controls for ollama to match the new format structure in the request json.
Fixes
Fixed the work flow process return display in the debug recorder.
Fixed all the process control displays in the debug recorder to display properly.
v0.4.2
Features
You can now run simple assistant prompts with the -t --test argument for the Dandy cli.