I tried switching to gemini-3-pro and am seeing this error: failed with status code: BadResponse 2\\nMessage: Failed to parse into a response accepted by baml_runtime::internal::llm_client::primitive::vertex::types::VertexResponse: {\\\"candidates\\\":[{\\\"content\\\":{},\\\"finishReason\\\":\\\"MALFORMED_FUNCTION_CALL\\\",\\\"finishMessage\\\":\\\"Malformed function call: call:default\\\\n{ .... "usageMetadata\\\":{\\\"promptTokenCount\\\":687936,\\\"totalTokenCount\\\":687936,\\\"trafficType\\\":\\\"ON_DEMAND\\\",\\\"promptTokensDetails\\\":[{\\\"modality\\\":\\\"TEXT\\\",\\\"tokenCount\\\":687936}]},\\\"modelVersion\\\":\\\"gemini-3-pro-preview\\\",\\\"createTime\\\":\\\"2026-01-27T21:18:35.108327Z\\\",\\\"responseId\\\":\\\"qyt5aafOBsSB6tkPtefKwAk\\\"}\\n\\nCaused by:\\n missing field `parts`\", It does not happen all the time but enough that I will have to revert the switch to Gemini

Hi all — I have a quick question regarding streaming behavior. I’m currently integrating BAML into a LangChain-compatible chat model (ChatBaml).

python
async def _astream()
  ...
  stream = self.b.stream.ChooseTool(baml_state, {"tb": tb}) # tb for @@dynamic usage in baml_src/
  async for partial in stream:
    ...

Right now partial.model_dump() is accumulated, e.g.:

- 1st emit: {"selected_tool": None}
- 2nd emit: {"selected_tool": {"name": None}}
- 3rd emit: {"selected_tool": {"name": "rep"}}
- 4th emit: {"selected_tool": {"name": "reply_to_user"}}

For LangChain streaming, I’d prefer delta-style updates (only the changed field/value per emit). Is there a configuration to make partial emit delta-style updates instead of accumulated state?

When using Kimi-K2.5 model via OpenAI-compatible API with BAML (Boundary ML) for structured output, the model fails to follow the specified JSON schema and instead outputs its internal tool-calling formats. Environment - Model: Kimi-K2.5 - API Endpoint: https://www.sophnet.com/api/open-apis/v1 - Provider: openai-generic

Hey Guys, Just started using BAML for the first time today so please forgive me if i'm missing something obvious, but when i click "Test {Fuction}" Directly in VSCode to test a function, i keeping getting an invalid auth error shown in the screen recording attached even though i have setup my .env and added it in BAML as well. It only works on the 2nd run of re-clicking run without the auth issue. https://cdn.discordapp.com/attachments/1467629807713124532/1467629808996450386/Screen_Recording_2026-02-01_at_1.07.14_PM.mov?ex=69811445&is=697fc2c5&hm=627e482386bd86758817e685c5fd09cce5c155f6bf92f31f8e3585d36dc8fbb2& https://cdn.discordapp.com/attachments/1467629807713124532/1467629809831252069/Screenshot_2026-02-01_at_1.14.25_PM.png?ex=69811445&is=697fc2c5&hm=727aba57f704b177fb3e65407d0b023c6bab5131d7dd86bad801ce2d9a3a8cba&

Hi BAML team! with a simple function like this

function Invoke() -> string {
    client ChatBaml
    prompt #"
        {{ _.role( 'user' ) }}
        {{ 'Hello, how are you?' }}
    "#
}

vllm renders it as ( Hermes format):

<|im_start|>user\nHello, how are you?<|im_end|>\n<|im_start|>assistant\n

Is there currently (or planned) support for controlling / injecting a custom string right at the start of the assistant turn, for example: {{ _.role( 'assistant', prefix='assistant{"name": “browser' ) }} will be rendered as

...<|im_end|>\n<|im_start|>assistant<tool_call>{"name": “browser

This pattern is very useful for reliable tool-calling agents (inspired by the Manus blog post "Context Engineering for AI Agents: Lessons from Building Manus"). Thanks in advance!

Hi, I have a code as below:

from baml_py import Image  

img = Image.from_url(f.absolute().as_posix())  
content = b.DescribeMedia(img)

When i read the image i get an error at `content = b.DescribeMedia(img)`:

baml_py.internal_monkeypatch.BamlValidationError: BamlValidationError(message=Internal Failure: Failed to fetch media: builder error

Caused by:
    relative URL without a base - Failed to fetch media: builder error

Caused by:
    relative URL without a base, raw_output=, prompt=N/A, detailed_message=Parsing error: Internal Failure: Failed to fetch media: builder error

Caused by:
    relative URL without a base - Failed to fetch media: builder error

Caused by:
    relative URL without a base
Prompt: N/A
Raw Response: )

It is an absolute path to the file. The test case below works fine :

test Test {
  functions [DescribeMedia]
  args {
    img {
      file "<ABSOLUTE_PATH>"
    }
  }
}

I'm building a multi-turn agent loop where the BAML function is called repeatedly until a final answer is produced. The conversation history (including past tool calls and their results) is passed as a

messages

parameter and rendered in the prompt template.

Sup BAML! Usually this is Anders but I've been tasked with implementing all our LLM providers so I've been working heavily with the BAML client registry 🙂 I'm trying to make it so we can leverage the OpenAI ChatGPT Plus/Pro subscription in our TUI. I'm getting an oauth token from OpenAI. Then I need to call the codex model specifically using the Responses API. I'm trying to do it like so:

const bamlProvider = (providerId === 'openai' && auth?.type === 'oauth')
    ? 'openai-responses' as const
    : def.bamlProvider

  // The options passed to BAML:
  cr.addLlmClient('ChatClientNoRetryAnthropicOnly', 'openai-responses', {
    model: modelId,
    api_key: auth.accessToken,
    base_url: 'https://chatgpt.com/backend-api/codex',
    headers: {
      'ChatGPT-Account-Id': auth.accountId,
    },
  })

But I keep getting this error back from OpenAI:

Request failed with status code: 400 Bad Request, 
{"detail":"Instructions are required"}

I thought the BAML openai-response provider would properly map system instructions to the right parameters. But is something different with the codex endpoint? Do I need to use the Modular API to manually construct the proper request here?

Never seen this one in previous installs - Title covers the issue. Not sure what I'm missing. https://cdn.discordapp.com/attachments/1472416235852664887/1472416242869604443/image.png?ex=69927dfc&is=69912c7c&hm=2ca8b8317462b9168a3c323cd89a4eb03371fd1762de95aeca158e26a6002c27&

I'm having a lot of trouble getting the prompt optimizer to do anything. It seems like it just runs through the tests and gives me a "complete!". It doesn't look like it's doing any actual updating of the prompt. I did change the gepa client to be Gemini 2.5 pro in vertex (my company can only use models in vertex). I've got a bunch of tests with assert statements, and I even intentionally gave a really bare bones prompt (accuracy only 34%) and it still didn't make any updates to it

I’m working on a use case involving structured extraction from Excel files that contain 200 to 300 rows across multiple sheets. Currently, I’m chunking the data and sending it piece-by-piece to OpenAI to get structured output, then combining the results. However, I want to explore whether BAML can handle this extraction in one go, since sending the entire file at once to OpenAI isn’t feasible. Additionally, the next step in the process involves a mapping agent that takes the extracted content from the Excel file, maps it to specific database attributes, and then generates SQL queries to load the data essentially an ETL workflow. I’m interested in hearing if anyone is working on a similar use case or has experience using BAML for this kind of structured data extraction and mapping from Excel. I’d appreciate any insights, ideas, or advice on how to approach this efficiently, especially regarding the capabilities and limitations of BAML in this context. If you’re open to discussing or collaborating, I’m eager to connect and learn more.

Hi, I'm currently exploring using BAML for a new project at work. I can't seem to get Entra ID auth to work. Is this not supported yet? In the meantime I use the modular API to use my own python client from the openAI SDK. That also has the plus of utilizing responses API.

Hi BAML! I have a LLM response that is encountering a BAML coerce error. My guess would be the triple backticks with brackets causing the error. (I'll include the LLM response in the thread, couldnt fit it in this message) I'm not sure if this is a bug on BAML's end or if i'm not using BAML correctly. Appreciate any help/advice. Thank you! Truncated error message to fit in this message

BAML COERCE ERROR:
Failed to coerce value: ParsingError { scope: [], reason: "Failed to find any ExtractedLiterals @stream.not_null in 4 items", causes: [ParsingError { scope: [], reason: "Failed to find any ExtractedLiterals @stream.not_null in 3 items", causes: [ParsingError { scope: [], reason: "Failed while parsing required fields: missing=1, unparsed=0", causes: [ParsingError { scope: [], reason: "Missing required field: explanation", causes: [] }] },
...
 causes: [] }] }] }, ParsingError { scope: [], reason: "Failed while parsing required fields: missing=1, unparsed=0", causes: [ParsingError { scope: [], reason: "Missing required field: explanation", causes: [] }] }] }

Any ideas on why that fails?

A retry_policy fires on network issues; can we set something up to have it retry on BamlValidationErrors too? If not, what's an elegant workaround?

Even though the raw LLM response appears valid and follows the required schema, the parser returns an incorrect result. Parsed response

{
  "selected_tool": {
    "name": "reply_to_user",
    "arguments": {
      "role": "assistant",
      "content": "string" // expected full extracted content "Jakob :four_leaf_clover:..."
    }
  }
}

Raw LLM Response

... (reasoning omitted)

Final check of the schema requirement: "Answer in JSON-like string (no double quotes in keys) using this schema."
Okay, I will construct the JSON object.

Content string:
Jakob :four_leaf_clover: \u0000
@jcsrb
Cứ 6 giờ debug sẽ giúp cho
bạn tiết kiệm được 5 phút đọc
document
@vuicoding

I'll put this into the `content` field.
</think>

{
  selected_tool: {
    name: "reply_to_user",
    arguments: {
      role: "assistant",
      content: "Jakob :four_leaf_clover: \\u0000\n@jcsrb\nCứ 6 giờ debug sẽ giúp cho\nbạn tiết kiệm được 5 phút đọc\ndocument\n@vuicoding"
    }
  }
}

https://cdn.discordapp.com/attachments/1493103826906124288/1493104266368389120/image.png?ex=69ddc135&is=69dc6fb5&hm=dfcfd504f9cab58b6c60a30c1aeb39ecd4195158654f33d46b8d019710efeb08&

I have this custom client definition: client Gemini3FlashPreviewVertex { provider vertex-ai retry_policy Exponential options { model "gemini-3-flash-preview" project_id "[my_project_name]" location "global" // project_id env.GCP_PROJECT_ID // location env.GCP_LOCATION generationConfig { thinkingConfig{ thinkingLevel "low" } } } } If I hard code the

project_id

and

location

, it works. If I use env lookup as commented out, it doesn't. my monorepo is set up as: project/ - backend/ - frontend/ BAML is under backend. I tried to put .env in the root as well as backend/ Neither approach worked. Any help on how to get this working is appreciated.

I have been using the 0.221.0 version of baml but it is no longer available on https://github.com/BoundaryML/baml/releases - what happened?

Attempting to create an agent that uses an existing test process document (future scope - vector database) to perform guided form completion. Taking inspiration from https://github.com/BoundaryML/baml-examples/blob/main/form-filler/baml_src/form_filler.baml - What is the best method to go about setting this up under BAML and Python?

Seems using the default "openai-responses/gpt-5-mini" client from resume.baml and executing a test from the baml file leads to

InvalidAuthentication (401) Request failed with status code: 401 Unauthorized, 
{
  "error": {
    "message": "Incorrect API key provided: $OPENAI_***_KEY. You can find your API key at https://platform.openai.com/account/api-keys.",
    "type": "invalid_request_error",
    "param": null,
    "code": "invalid_api_key"
  }
}

Request options: {"model":"gpt-5-mini"}

Based on preliminary testing, changing to a named client from clients.baml (i.e. CustomGPT5Mini) fixes this behavior. You can also just rerun the test from the Playground once it loads to get it working with no changes required. Not sure if this is expected behavior or known bug, figured I'd leave it here anyway. https://cdn.discordapp.com/attachments/1504994600941785100/1504994601109426306/image.png?ex=6a0902f2&is=6a07b172&hm=a7a3506b9f00d2827300170bafed70d9dcd6193055237c50ab6103f5461daf91&

Hey, I've been looking around but was wondering if there was an integration with BAML such that it handles the explicit context caching that gemini supports, per https://ai.google.dev/gemini-api/docs/caching#explicit-caching I know anthropic is supported with the

cache_control

blocks, but I didn't see anything in the docs for gemini.

Hi! I’m leading a small but growing business team, and we’re looking for a dedicated developer who wants to make a real impact and grow with us. What we need: - 2+ years of professional software development experience - Solid understanding of web development (frontend and/or backend) - Reliable, proactive, and ready to contribute to a small team Why join us: - Flexible remote work from US, EU, or LATAM - Be part of a collaborative, supportive team - Opportunities to take ownership and grow professionally If you’re ready to contribute, learn, and make an impact, we’d love to hear from you!

GenerateCode(input: GenerateCodeRequest, conversation_history: Message[]) -> GenerateCodeResponse | AskUserQuestion I saw the example in the doc but it a single schema which has the response and tool calls. I want to keep actual response and tool calls separate. Like GenerateCodeResponse | AskUserQuestion In general, the docs are not fully reflecting the Golang SDK

Hi, I am building a system that allows users to navigate a Graphdatabase and financial records, with the goal to write reports, request missing data, or explore the data. Eventually with the goal to use a2ui or something similar to dynamically create visuals. Which means I will create my own CLI to expose all the functions. And of course will use BAML, currently canary version, to drive the CLI. Is there a guide, read up resource, on best practices or better yet don't s for writing custom harnesses for custom tools? Especially as it is non coding. And yes I read the Claude code leak and the cool tools that visualize how they work etc. But was wondering if I am missing something.

LLM response is right and also the baml parsing is also right. Just that in studio its showing null. https://cdn.discordapp.com/attachments/1525107289521000479/1525107289659408395/Screenshot_2026-07-10_at_5.21.28_PM.png?ex=6a522e58&is=6a50dcd8&hm=5d681a07ad5271f5b4490f469a71a95fe3bf2df408cb593499e8cd42bfe58fd6&

I am trying to extract the model used in the baml function call. I see no straightforward way.

goroutine 473 [running]:
reflect.Value.Elem({0x2383220?, 0xc00103f530?, 0xc00103f380?})
    /usr/local/go/src/reflect/value.go:1265 +0x190
github.com/boundaryml/baml/engine/language_client_go/baml_go/serde.decodeListValue(0xc000b15ae0, {0x13?, 0xc00103f380?})
    /root/.cache/go-build/github.com/boundaryml/baml@v0.222.0/engine/language_client_go/baml_go/serde/decode.go:114 +0x2b3
github.com/boundaryml/baml/engine/language_client_go/baml_go/serde.Decode(0xc00103dc80, {0xd0?, 0xc00103f380?})
    /root/.cache/go-build/github.com/boundaryml/baml@v0.222.0/engine/language_client_go/baml_go/serde/decode.go:581 +0x14b
github.com/boundaryml/baml/engine/language_client_go/baml_go/serde.decodeMapValue(0xc000b15a90, {0x13?, 0xc00103f380?})
    /root/.cache/go-build/github.com/boundaryml/baml@v0.222.0/engine/language_client_go/baml_go/serde/decode.go:145 +0x2a5
github.com/boundaryml/baml/engine/language_client_go/baml_go/serde.Decode(0xc00103dac0, {0x60?, 0xc00103f380?})
    /root/.cache/go-build/github.com/boundaryml/baml@v0.222.0/engine/language_client_go/baml_go/serde/decode.go:583 +0x4af
github.com/boundaryml/baml/engine/language_client_go/baml_go/serde.decodeListValue(0xc000b15900, {0x13?, 0xc00103f380?})
    /language_client_go/baml_go/raw_objects/utils.go:231 +0x20c
    /language_client_go/baml_go/serde/decode.go:123 +0x185
github.com/boundaryml/baml/engine/language_client_go/baml_go/serde.Decode(0xc00103d900, {0xb0?, 0xc00103f380?})
    language_client_go/baml_go/serde/decode.go:581 +0x14b
github.com/boundaryml/baml/engine/language_client_go/baml_go/serde.decodeMapValue(0xc000b155e0, {0x13?, 0xc00103f380?})
    /root/.cache/go-build/github.com/boundaryml/baml@v0.222.0/engine/language_client_go/baml_go/serde/decode.go:145 +0x2a5
github.com/boundaryml/baml/engine/language_client_go/baml_go/serde.Decode(0xc00103d440, {0x80?, 0xc00103f380?})

https://cdn.discordapp.com/attachments/1527014242811908231/1527014243222818927/Screenshot_2026-07-15_at_11.33.50_PM.png?ex=6a591e56&is=6a57ccd6&hm=ffcacc4d390b578453588e52e1922c84a7b740f04b9e126d4527ca1add5339d7&

Hello there, running the bamlcode repo with the BrainOpenAI client, and I keep running into rate limit issues with OpenAI. What's the remedy here? https://cdn.discordapp.com/attachments/1532470005562868005/1532470005961330831/image.png?ex=6a6cf769&is=6a6ba5e9&hm=1911f6fb99f17318608c9e179a3f1da25ffd0ceacaa02915adbf6bfc1a0c8d4c&

So, latest rust-sdk on crates.io is 0.221.0 But latest baml-cli is 0.225.0 How to use baml in rust then? baml-cli generated code still imports

baml::

BAML Version:

BAML 0.15.1-nightly

Situation: Agent loop ran

baml test

(120 tests, 0s) repeatedly, creating lots of .bamlprof files, adding >1GB of files in 3.5h. OS X Spotlight indexing then froze the machine trying to pick it up. Any way to disable or cap profiles, or suggestions on ways to mitigate this?

I am receiving an error running github dependabot. It looks like v0.226.1 was re-tagged. Source, Commit, Tag time, zip hash proxy.golang.org, 044992b, 2026-08-19 03:21:21Z, h1:MSABQSEYQjmFIORLT5OXOUqjbQjTxErU+T950EPAc4Y= github.com today, 70612e5, 2026-08-19 04:30:31Z, h1:Jr0Vpdquv3tVSWoKFyja8NDWUndGDWtfHSC5vJBPxTs= Can you please publish a new v0.226.2 with the latest?