Ty/ollama local models

[AlephNotation]

adds ollama models

[matildepark]

If we can't get tiny models to .browse(), do we feel ok about them doing .get() and .do()? It would be nice to incorporate this, even with limitations.

[matildepark]

Current status (over phone):

• works okay with page.get(), not so sure about anything else.
• Test, document, merge into dev.

[matildepark]

Been doing some experiments with examples/findEmail.ts and node-llama-cpp, and I needed to set the context and batch size in generateObjectLocal.ts:

  const context = new LlamaContext({ model, contextSize: 32168, batchSize: 32168 });

Otherwise, it would throw:

/Users/maru/git/hdr/nolita/node_modules/node-llama-cpp/llama/llama.cpp/src/llama.cpp:14466: GGML_ASSERT(n_tokens_all <= cparams.n_batch) failed

The model in question I picked was capybarahermes-2.5-mistral-7b. Once I set contextSize and batchSize, it did get the emails off the page. If you try to get it to browse, it won't generate a command.

  issues: [
    {
      code: 'too_small',
      minimum: 1,
      type: 'array',
      inclusive: true,
      exact: false,
      message: 'Array must contain at least 1 element(s)',
      path: [ 'command' ]
    }
  ],
  addIssue: [Function (anonymous)],
  addIssues: [Function (anonymous)],
  errors: [
    {
      code: 'too_small',
      minimum: 1,
      type: 'array',
      inclusive: true,
      exact: false,
      message: 'Array must contain at least 1 element(s)',
      path: [ 'command' ]
    }
  ]
}

[AlephNotation]

@matildepark good to go?