This project contains a virtualized development environment utilizing the anthropic Python SDK for testing and improving Claude's responses to a multitude of (predominantly scientific) queries across different model versions. It is designed with the VSCode IDE in mind and thereby includes various files related to installing extensions, configuring settings, etc. Feel free to clone this repo and run it in a separate IDE of your choice, deleting or ignoring irrelevant files.
We implement a Python class, LatexParser, capable of reading a full LaTeX token response from Claude into a .tex file before compiling it and creating a PDF that is easy to interpret.
Users can issue a new query to Claude by creating a file <choice_of_name>.py and copying the outline of the .py files in the src/interactive_model/queries path relative to the container's working directory. Once created, users should configure the ClaudeModel class to take values for the claude_model, max_tokens, and system_prompt parameters. See the ClaudeModel docstring for an outline of expected values. Next, simply run the <choice_of_name>.py file containing the custom query to Claude, passed as the query argument to the _query_claude() func. The project code will automagically read Claude's response and return a neatly compiled set of files <choice_of_name>.tex and <choice_of_name>.pdf in the src/latex path.
See the set of hello_world_latex.py/.tex/.pdf and superstring_theory.py/.tex/.pdf example files found in both of the aforementioned directories for concrete reference.
Currently, caution must be taken when querying Claude. One must carefully engineer both the query and Claude's system_prompt parameter to ensure any implementation of the LatexParser function receives a response from Claude that is entirely LaTeX. See again the src/interative_model/queries/hello_world_latex.py and src/interative_model/queries/superstring_theory.py files for an example of a definitive system prompt. An improvement to this project will soon be made to simply pick the LaTeX portion of a response from Claude. This will remove the LaTeX-only response requirement.
A last possible issue is the max_tokens parameter. Each of Claude's responses are quantified in terms of the number of "tokens" he produces to adeuquately respond to a given query. If you receive an error indicating that the max_tokens value was too small, thereby not allowing Claude to fully print his response to you, simply increase the number. In general, more complex queries result in longer and more complex respones, driving the number of output tokens upwards. A good tactic to fully avoid this problem is to just max out the parameter. Set max_tokens=10000.
Before executing any queries to Claude, the user must first set set their environment variable ANTHROPIC_API_KEY to an active Anthropic API key.
Simply running ANTHROPIC_API_KEY=<your-api-key> inside the Linux container's command line will update the value appropriately. Contact the project owner if you would like to try this out temporarily with a custom API key and he will create one for you.