How to use Large Language Models (LLMs) with OAK

Large Language Models (LLMs) such as ChatGPT have powerful text pattern matching and processing abilities, and general question answering capabilities. LLMs can be used in conjunction with ontologies for a number of tasks, including:

Summarizing lists of ontology terms and other OAK outputs
Annotating text using ontology terms
Reviewing ontology branches or different kinds of ontology axioms

For more on LLMs, see:

Google LLM guide

This guide is in 3 sections:

Summary of ontology-LLM tools that directly leverage OAK, but are not part of the OAK framework
How to use OAK in conjunction with existing generic LLM tools
The OAK LLM implementation

LLM frameworks that use OAK

OntoGPT

OntoGPT extracts knowledge from text according to a LinkML schema and LinkML dynamic value set specifications. OAK is used for grounding ontology terms.

CurateGPT

CurateGPT is a general purpose knowledge management and editing tool that uses LLMs for enhanced search and autosuggestions.

Talisman

Talisman allows for an LLM analog of the OAK enrichment command. It summarizes collections of terms or descriptions of genes.

Using OAK in conjunction with existing LLM tools

LLMs such as ChatGPT can take any kind of textual output, including outputs of OAK.

For example, you could query all T-cell types:

runoak -i sqlite:obo:cl labels .descendant//p=i "T cell"

And then copy the results into the ChatGPT window and ask “give me detailed descriptions of these T-cell types”.

This kind of workflow is not very automatable. OAK is designed in part for the Command Line, so LLM CLI tools such as the datasette llm tool pair naturally

pipx install llm
runoak -i sqlite:obo:cl labels .descendant//p=i "T cell" | llm --system "summarize the following terms"

OAK LLM Adapter

Annotation

runoak -i llm:sqlite:obo:hp annotate "abnormalities were found in the eye and the liver"

Suggesting Definitions

runoak -i llm:sqlite:obo:uberon generate-definitions \
     finger toe \
     --style-hints "write definitions in formal genus-differentia form"

Validating Definitions

The LLM adapter currently interprets validate-definitions as comparing the specified definition against the abstracts of papers cited in the definition provenance, or by comparing the definition against the database objects that are cited as definition provenance.

Here is an example of validating definitions for GO terms:

runoak --stacktrace -i llm:sqlite:obo:go validate-definitions \
  i^GO: -o out.jsonl -O jsonl

The semsql version of GO has other ontologies merged in, so the i^GO: query only validates against actual GO terms.

You can also pass in a configuration object. This should conform to the Validation Data Model

For example, this configuration yaml provides a specific prompt and also a URL for documentation aimed at ontology developers.

prompt_info: Please also use the following GO guidelines
documentation_objects:
  - https://wiki.geneontology.org/Guidelines_for_GO_textual_definitions

All specified URLs are downloaded and converted to text and included in the prompt.

The configuration yaml is passed in as follows:

runoak --stacktrace  -i llm:{claude-3-opus}:sqlite:obo:go validate-definitions \
     -C src/oaklib/conf/go-definition-validation-llm-config.yaml i^GO: -O yaml

Validating Mappings

The LLM adapter validates mappings by looking up info on the mapped entity and comparing it with the main entity.

runoak --stacktrace -i llm:{gpt-4}:sqlite:obo:go validate-mappings \
   .desc//p=i molecular_function -o out.jsonl -O jsonl

Selecting alternative models

If you are using the LLM Adapter then by default it will use a model such as gpt-4 or gpt-4-turbo (this may change in the future).

You can specify different models by using the {} syntax:

runoak -i llm:{gpt-3.5-turbo}:sqlite:obo:cl generate-definitions .descendant//p=i "T cell"

We are using Datasette LLM package which provides a plugin mechanism for adding new models. See Plugin index.

However, LLM can sometimes be slow to add new models, so here it can be useful to the awesome LiteLLM package, which provides a proxy to a wide range of models.

Installing LLM plugins

The llm command line tool makes it easy to access other models via its extensible plugin system.

Normally, you would do something like this:

pipx install llm
llm install llm-gemini
llm -m gemini-pro "what is the best ontology?"

However, this will install the plugin in a different environment from OAK. If you are running OAK as a developer, then you can do this:

cd ontology-access-kit
poetry run llm install llm-gemini

This will install the plugin in the same environment as OAK.

If you need to update this:

cd ontology-access-kit
poetry run llm install -U llm-gemini

TODO: instructions for non-developers.

Mixtral via Ollama and LiteLLM

ollama run mixtral

pipx install litellm
litellm -m ollama/mixtral

Next edit your extra-openai-models.yaml as detailed in the llm [other model docs](https://llm.datasette.io/en/stable/other-models.html):

- model_name: ollama/mixtral
  model_id: litellm-mixtral
  api_base: "http://0.0.0.0:8000"

Then you can use the model in OAK:

runoak -i llm:{litellm-mixtral}:sqlite:obo:cl generate-definitions .descendant//p=i "T cell"

Mixtral via groq and LiteLLM

groq provides an API over souped-up hardware running Llama2 and Mixtral. You can configure in a similar way to ollama above, but here we are proxying to a remote server:

. code-block:: bash

pipx install litellm litellm -m groq/mixtral-8x7b-32768

Next edit your extra-openai-models.yaml as detailed in the llm [other model docs](https://llm.datasette.io/en/stable/other-models.html):

- model_name: litellm-groq-mixtral
  model_id: litellm-groq-mixtral
  api_base: "http://0.0.0.0:8000"

Then you can use the model in OAK:

runoak -i llm:{litellm-groq-mixtral}:sqlite:obo:cl validate-mappings .descendant//p=i "T cell"