Platform Assistant
Otoshek includes a real AI chat assistant, not just a placeholder chat button. It is already wired into the platform shell with a floating modal, a dedicated chat page, live WebSocket streaming, knowledge-base retrieval, and scope guardrails that keep answers focused on Otoshek.
Feature snapshot
This is what the platform assistant currently ships with in the live platform implementation.
| Capability | What ships today |
|---|---|
| Access points | Floating Ask AI modal on most pages plus a dedicated /chat page |
| Frontend UI | assistant-ui thread with starter prompts, welcome copy, and streamed replies |
| Backend engine | LangGraph assistant graph running through Django Channels |
| Transport | WebSocket stream at /ws/platform_assistant/ with run-scoped events |
| Knowledge sources | Product overview plus deeper platform knowledge-base documents loaded from local files or blob storage |
| Reasoning UX | Short progress labels are shown separately from the final answer text |
| Memory model | Thread memory exists for the current session but is not durable long term |
| Scope guardrail | The assistant stays focused on Otoshek and redirects off-topic questions |
How the user experience works
The assistant is designed to feel available everywhere without forcing a single chat layout onto every screen size.
Ask AI follows users around the platform
The main shell mounts a floating assistant on non-chat pages, so users can open help without leaving the workflow they are already in.
Small screens switch to the full chat page
On mobile and smaller layouts, opening the assistant routes users to the dedicated /chat experience instead of keeping a cramped modal open.
Desktop state is remembered
The assistant stores open and closed state in local storage and gently auto-opens on first desktop visits after a short delay.
Users get starter guidance immediately
The thread includes a welcome message and built-in starter prompts like platform overview and ideal users, which helps first-time visitors ask better questions.
How answers are generated
Under the UI, the assistant runs as a small LangGraph workflow rather than a single blind model call.
- Each message is normalized and added to the current thread history before any answer logic runs.
- A first router classifies the request as a greeting, an Otoshek question, a conversation follow-up, or out of scope.
- Platform questions take a second route: answer from the overview when possible, or retrieve deeper documentation chunks when needed.
- Deeper questions use a simple retrieve-once, answer-once RAG path instead of a heavier multi-pass agent loop.
- The backend streams
messages,custom, andupdatesevents, and the frontend turns those into live text plus short reasoning/progress states.
What the assistant knows
The assistant is grounded in Otoshek documentation instead of acting like a generic free-form chatbot.
Knowledge sources
These are the documents and retrieval layers the current assistant uses to answer questions.
- Fast-path answers come from
platform_assistant/platform_docs/overview.md. - Deeper answers come from
platform_assistant/platform_docs/platform_knowledge_base.txtsplit into retrievable chunks. - The loader can read those documents from local files by default or from blob storage when document storage variables are configured.
- If retrieval is weak or the model backend is unavailable, the assistant is designed to answer honestly instead of bluffing.
Guardrails
The assistant is intentionally narrow so it can stay useful and trustworthy.
The first routing layer explicitly separates Otoshek questions from greetings, follow-ups, and out-of-scope requests. Off-topic requests are redirected back toward Otoshek instead of treated like a general-purpose chatbot prompt.
When the overview or retrieved documentation does not support an answer confidently, the assistant is designed to say so rather than inventing details.
Current limitations
The assistant is already useful, but it is intentionally simple in a few important places.
- Memory uses LangGraph
InMemorySaver, so it is useful for the active session but not durable across process restarts. - Thread identity is currently derived from the live socket, which means reconnects and refreshes do not preserve the same durable conversation id.
- Follow-up questions currently answer from prior thread context only and do not yet decide when to trigger fresh retrieval automatically.
- The UI does not yet render source citations, even though the backend preserves selected document metadata internally.
Reference app note
This feature is not just described in the docs. The current Tailwind TypeScript reference app includes the same platform assistant backend module, chat page, modal wrapper, runtime hook, and WebSocket route.
