v0.5.0 - Files

What's Changed

New features

• Files API

  You can now upload files to the api. Then instead of resending an image with each request, just pass a file reference in your prompt:

  f = genai.upload_file(path="image.png")
  m = genai.GenerativeModel(model_name=...)
  response = m.generate_content(["Please describe this file:", f])
  

  • Add initial prototype supporting the files API by @MarkDaoust in #249
  • Files API: improve error messages, add delete_file, add .uri property by @TYMichaelChen in #260

• Semantic retriever

  Using answer.generate_answer, you can now either pass "inline_passages" to include the reference material in the request:

  from google.generativeai import answer
  answer.generate_answer(
      content=question,
      inline_passages=splitter.split(document)
  )
  

  Or pass a reference to a retriever Document or Corpus:

  from google.generativeai import answer
  from google.generativeai import retriever
  my_corpus = retriever.get_corpus('my_corpus')
  genai.generate_answer(
      content=question,
      semantic_retreiver=my_corpus
  )
  

  • Finish Semantic Retriever. by @shilpakancharla in #230

• System instructions

  When creating a model you can pass a string (or Content) as system_instructions: genai.GenerativeModel(system_instructions="Be good!")

  • Add basic SI and tool_config support by @MarkDaoust in #257
  • Add system_instruction to model repr by @markmcd in #272

• Function Calling - Tool Config

  The list of tools available for function calling during a chat session is typically constant. The new tool_config argument lets you switch the function calling mode between None (No function calls), Auto (The model chooses to return text or a FunctionCall), and Any (the model must call a function).

  model = genai.GenerativeModel(tools=[my_functions])
  chat = model.start_chat()
  chat.send_message(tool_config={"function_calling_config": "ANY"})
  

  In Any mode the model will return a function call. You may also pass allowed_function_names to restrict the calls to a subset of the available functions: `chat.send_message(tool_config={"function_calling_config": "ANY", "allowed_function_names": [...]})

  In Any mode the model can use constrained decoding to more strictly follow the argument specifications in the function declarations.

  • Add basic SI and tool_config support by @MarkDaoust in #257
  • Function calling mode patch by @mayureshagashe2105 in #271

Docs, bug fixes, and minor updates.

• Update README.md by @aertoria in #256
• Fix incorrect reference to self.candidates[0].parts, self.candidates[0].parts[0].text by @Andy963 in #239
• Fix TypedDict usage for compatibility with PyDantic by @MarkDaoust in #259
• Simplify the Python SDK README by @random-forests in #267
• Sdk update to use glm 0.6.1 by @mayureshagashe2105 in #270

New Contributors

• @aertoria made their first contribution in #256
• @TYMichaelChen made their first contribution in #260
• @random-forests made their first contribution in #267

Full Changelog: v0.4.1...v0.5.0