[Bug]: GoogleGenAI raises error when using stream_complete

[yeosadchuk]

Bug Description

When using GoogleGenAI llm through the API, the stream_complete is not working and raises an error.

"pydantic_core._pydantic_core.ValidationError: 1 validation error for ChatMessage
blocks.0
Input should be a valid dictionary or object to extract fields from [type=model_attributes_type, input_value='What is the capital of France? answer short.', input_type=str]
For further information visit https://errors.pydantic.dev/2.12/v/model_attributes_type"

Version

0.14.7

Steps to Reproduce

from llama_index.llms.google_genai import GoogleGenAI

import config

llm = GoogleGenAI(
model="gemini-2.0-flash",
api_key=config.GEMINI_API_KEY,
context_window=10000,
)

resp = llm.stream_complete(["What is the capital of France? answer short."])

Relevant Logs/Tracbacks

venv/lib/python3.11/site-packages/pydantic/main.py", line 250, in __init__
    validated_self = self.__pydantic_validator__.validate_python(data, self_instance=self)
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
pydantic_core._pydantic_core.ValidationError: 1 validation error for ChatMessage
blocks.0
  Input should be a valid dictionary or object to extract fields from [type=model_attributes_type, input_value='What is the capital of France? answer short.', input_type=str]
    For further information visit https://errors.pydantic.dev/2.12/v/model_attributes_type

[dosubot]

Hi @yeosadchuk! I'm Dosu and I’m helping the llama_index team.

The error happens because stream_complete expects structured message objects (like ChatMessage dicts), not plain strings. Recent changes to the GoogleGenAI integration enforce that the input must be a dictionary or object with extractable fields, matching the expected schema for message blocks. To fix this, pass a properly structured message object instead of a string. For example, you should use something like:

from llama_index.core.llms import ChatMessage, MessageRole

messages = [ChatMessage(role=MessageRole.USER, content="What is the capital of France? answer short.")]
resp = llm.stream_complete(messages)

This matches the internal validation and avoids the Pydantic error. The recent PRs clarify that message blocks are expected to be objects, not raw strings, which aligns with the error you're seeing (source, source).

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[yeosadchuk]

@dosu

The issue still persists. The error directs to a pydantic_core. ValidationError.

pydantic_core._pydantic_core.ValidationError: 1 validation error for ChatMessage
blocks.0
Unable to extract tag using discriminator 'block_type' [type=union_tag_not_found, input_value=ChatMessage(role=<Message...rance? answer short.')]), input_type=ChatMessage]
For further information visit https://errors.pydantic.dev/2.12/v/union_tag_not_found

[dosubot]

The error is happening because the ChatMessage object's blocks field must contain valid ContentBlock objects, each with a block_type discriminator (like "text" or "image") for Pydantic to validate them correctly. If you pass a string, a ChatMessage, or a dict that doesn't have block_type, you'll get the union_tag_not_found error.

To fix this, make sure you construct your messages like this:

from llama_index.core.llms import ChatMessage, MessageRole, TextBlock

messages = [
    ChatMessage(
        role=MessageRole.USER,
        blocks=[TextBlock(text="What is the capital of France? answer short.")]
    )
]
resp = llm.stream_complete(messages)

You can also use the content argument for backward compatibility, which will automatically wrap your string in a TextBlock:

messages = [
    ChatMessage(role=MessageRole.USER, content="What is the capital of France? answer short.")
]
resp = llm.stream_complete(messages)

Both approaches ensure the blocks field is populated with a valid ContentBlock (with block_type="text"), which resolves the validation error. More details on the expected schema and the role of block_type are in the codebase and recent PRs (source).

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[yeosadchuk]

@dosu

I am following your advice. However, the issue persists. The same error appear.

The code I run:
messages = [ ChatMessage( role=MessageRole.USER, blocks=[TextBlock(text="What is the capital of France? answer short.")], ) ] resp = llm.stream_complete(messages)

The error I get:

pydantic_core._pydantic_core.ValidationError: 1 validation error for ChatMessage
blocks.0
Unable to extract tag using discriminator 'block_type' [type=union_tag_not_found, input_value=ChatMessage(role=<Message...rance? answer short.')]), input_type=ChatMessage]
For further information visit https://errors.pydantic.dev/2.12/v/union_tag_not_found

[dosubot]

This error is likely due to a mismatch between the expected input format for stream_complete and how the message objects are being constructed or passed in your current version. In LlamaIndex, each block in a ChatMessage must be an instance of a block class (like TextBlock) that has a block_type field set, and the list of blocks should not contain nested ChatMessage objects or other types without block_type—otherwise, Pydantic can't validate them and you'll see the union_tag_not_found error (reference).

In some cases, especially with certain versions, you may need to serialize your message objects before passing them to the LLM. Try converting your messages to dictionaries using .model_dump():

messages = [
    ChatMessage(
        role=MessageRole.USER,
        blocks=[TextBlock(text="What is the capital of France? answer short.")]
    )
]
resp = llm.stream_complete([msg.model_dump() for msg in messages])

If this still doesn't work, there may be a version-specific bug or a breaking change in how message schemas are validated (see related issue). Recent PRs have addressed similar problems by ensuring blocks are explicitly set as lists of block instances, not appended or nested (see PR). If you're on 0.14.7 and still seeing this, it may be worth checking if upgrading to the latest release resolves the issue.

Let me know if serializing with .model_dump() works or if you see a different error!

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[lf2foce]

The output was streaming correctly on llama-index-llms-google-genai==0.6.2, but after upgrading to 0.7.1, streaming no longer works.

[yeosadchuk]

@dosu

The same error persists with a new serializing. I believe it`s a validation error in the llamaindex core code.

What is the latests version if not the 0.14.7? This is not helpful.

[dosubot]

The latest officially released version of llama-index-core is 0.14.6 as of 2025-10-26. If you are using 0.14.7, it may be a pre-release or a typo, since 0.14.6 is the most recent version listed in the release notes I found. For the GoogleGenAI integration, the latest version is 0.7.1, which includes recent fixes for streaming and message block handling. Make sure your core and integration package versions are compatible, as mismatches can cause validation errors like the one you're seeing.
See release details

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[logan-markewich]

The full traceback is going to be the most helpful here. I'll try to repro