mirror of
https://github.com/hanzoai/insights.git
synced 2026-11-02 14:32:45 +00:00
zeekay
b97cdc1087
Eleven products ship a `max_tools.py`, and every one of them was dead: they subclass `MaxTool`, which was defined in the upstream `ee/` tree this deployment does not carry. Nothing imported them, so nothing said so. `MaxTool` is written natively here. It is a plain class rather than a LangChain `BaseTool` because the assistant's loop speaks the OpenAI dialect directly, so what a tool owes it is a JSON schema and a callable; the graph scaffolding `BaseTool` exists to serve has nothing standing on it. Its surface is what the tool files actually use and nothing more: `_team`, `_user`, `context`, `_config`, `name`/`description`/`args_schema`/`context_prompt_template`, one of `_run_impl` / `_arun_impl`, and the access and dangerous-operation hooks. Eight products are ported. Three — replay, revenue_analytics and web_analytics — also want `TaxonomyAgentToolkit`/`TaxonomyAgentNode`/`base_final_answer`, a graph the upstream is itself migrating off, and are left alone. The registry names them, so porting one is a visible change rather than a silent one. Tenancy is structural, not documentary. `team` is a required keyword argument from the authenticated request, arguments are validated through `args_schema` before a tool sees them, and the registry refuses to expose any tool whose schema declares a team, project or organization. No tool needed that fix today; the guard is there so none can acquire it later. Two things were broken beyond the missing base class, both from the debrand: cdp imported `compile_iql`, which nothing defines — the function is `compile_script` — and its test still spoke of `<iql_code>` where the tool reads `<script_code>`. user_interviews read `settings.OPENAI_BASE_URL`, which is not a setting here, and called an API the gateway does not serve; it now reaches a model the one way this product does.
78 lines
2.8 KiB
Python
78 lines
2.8 KiB
Python
"""The chat model a product tool reaches for, and the error it raises on bad output.
|
|
|
|
Tools that ask a model for structured output use `MaxChatOpenAI`. It is a
|
|
`ChatOpenAI` pointed at the Hanzo AI gateway with this deployment's own IAM
|
|
bearer — the same seam `assistant.py` uses, so there is one way to reach a model
|
|
from this product and not a second one.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
from typing import Any
|
|
|
|
from django.conf import settings
|
|
|
|
from langchain_core.messages import BaseMessage, SystemMessage
|
|
from langchain_openai import ChatOpenAI
|
|
|
|
from insights import iam
|
|
|
|
|
|
class PydanticOutputParserException(ValueError):
|
|
"""A model wrote something the schema would not accept.
|
|
|
|
Carries both halves because the tools that raise it feed them back into the
|
|
next attempt: the output so the model can see what it wrote, the message so
|
|
it can see what was wrong with it.
|
|
"""
|
|
|
|
def __init__(self, llm_output: str, validation_message: str):
|
|
super().__init__(f"{validation_message}, occurred at:\n```\n{llm_output}\n```")
|
|
self.llm_output = llm_output
|
|
self.validation_message = validation_message
|
|
|
|
|
|
_CONTEXT_TEMPLATE = """You are in the project "{project}" of the {organization} organization.
|
|
The project's timezone is {timezone}."""
|
|
|
|
|
|
class MaxChatOpenAI(ChatOpenAI):
|
|
"""A chat model on the Hanzo gateway, told which project it is working in.
|
|
|
|
`user` and `team` say who the generation is for. `billable` records that the
|
|
call is chargeable work; the gateway meters by the IAM identity presenting
|
|
the token, so this marks intent rather than deciding the charge.
|
|
"""
|
|
|
|
user: Any = None
|
|
team: Any = None
|
|
billable: bool = False
|
|
inject_context: bool = True
|
|
|
|
def __init__(self, **kwargs):
|
|
# Raises IamUnavailable rather than sending an unauthenticated request.
|
|
kwargs.setdefault("base_url", f"{settings.HANZO_API_URL}/v1")
|
|
kwargs.setdefault("api_key", iam.service_token())
|
|
super().__init__(**kwargs)
|
|
|
|
def _with_context(self, messages: list[BaseMessage]) -> list[BaseMessage]:
|
|
"""The messages, prefixed with which project this is, when asked for."""
|
|
if not self.inject_context or self.team is None:
|
|
return messages
|
|
return [
|
|
SystemMessage(
|
|
content=_CONTEXT_TEMPLATE.format(
|
|
project=self.team.name,
|
|
organization=self.team.organization.name,
|
|
timezone=self.team.timezone,
|
|
)
|
|
),
|
|
*messages,
|
|
]
|
|
|
|
def _generate(self, messages, *args, **kwargs):
|
|
return super()._generate(self._with_context(messages), *args, **kwargs)
|
|
|
|
async def _agenerate(self, messages, *args, **kwargs):
|
|
return await super()._agenerate(self._with_context(messages), *args, **kwargs)
|