The graph state the project passes around

Every node in the project reads and writes one shared object: the graph state. It is a TypedDict called VideoAuditState, holding the video URL and the id and file path the indexer produces, the transcript and on-screen text, the violations found, any system errors, and the final verdict. The video is deliberate that it is a TypedDict, not a Pydantic model.

Last updated: 15 Oct, 2026 · LLMOps

The knowledge base prepared the rules. Before the nodes can run, they need a shared shape to pass work along in. That shape is the state.

Defining the video audit state · from the End to End Multimodal LLMOps Project video · 1:44:00 to 1:47:00

What the state holds

python

from typing import TypedDict, Optional

class ComplianceIssue(TypedDict):
    category: str          # e.g. FTC disclosure, misleading claim
    severity: str          # critical or warning
    description: str       # the specific detail of the violation
    timestamp: Optional[str]   # where in the video, if known

class VideoAuditState(TypedDict):
    video_url: str             # the request's input
    video_id: str              # the id Video Indexer assigns
    local_file_path: Optional[str]   # the downloaded file, while it exists
    transcript: str            # what is said, from Video Indexer
    ocr_text: str              # the text on screen
    video_metadata: dict
    compliance_results: list[ComplianceIssue]
    system_errors: list[str]   # downloads that failed, timeouts
    status: str                # pass or fail
    final_report: str

The violation shape is fixed deliberately: every issue must carry a category, a severity and a description, and the video notes this is what stops the model from returning free-form text the rest of the system cannot act on. The timestamp is optional, the one field that may be missing.

The VideoAuditState TypedDict as a list of fields, each labelled with the node that fills it: video_url comes from the request; transcript, ocr_text and video_metadata are filled by the index node; compliance_results, status and final_report by the auditor node; and system_errors, drawn in red, by any node when something fails. A note says it is a TypedDict, not Pydantic: a declared shape LangGraph passes and merges rather than a runtime check.

Why a TypedDict, not Pydantic

A TypedDict is a plain dict with a declared shape. LangGraph passes it from node to node, each node returning the keys it filled in, and the graph merges them.
It documents the contract without validating at runtime, which is what the video wants here: a clear shape the nodes agree on, not a validation layer.
The structured violation type is the part that is enforced, later, by asking the model for exactly that shape of JSON.

Watch out. A TypedDict does not check types at runtime; it is a hint. If a node writes a string where the state declares a list, nothing stops it until something downstream breaks. The enforcement that matters here is on the model's output, covered in the auditor lesson, not on the state.

Previous: The knowledge base: rules into Azure AI Search
Next: Node one: indexing the video
See also: State

Try it yourself

Name which node fills each field of VideoAuditState.
Explain why the timestamp is the only optional field in a ComplianceIssue.

PreviousThe knowledge base: rules into Azure AI Search

This is what real progress feels like.

The graph state the project passes around

What the state holds

Why a TypedDict, not Pydantic

Related