Most ERPNext teams that want an internal AI assistant end up copy-pasting record data into a public ChatGPT tab and hoping nobody pastes something they shouldn't. That is a data-governance problem and a productivity dead-end: the model has no idea what a Sales Invoice or a BOM is in your instance, it cannot see the document the user is looking at, and every question is tied to a single vendor's API, pricing, and privacy posture. ERPNext core ships no conversational AI surface at all, and the marketplace options that exist are typically single-provider wrappers with no context injection and no fallback when that one provider is rate-limited or down. Privacy-conscious shops in particular want to keep the option of running everything against a local Ollama model and never letting record data leave the network.

ECOSIRE builds a proper Frappe app (its own module, installed via bench get-app and bench install-app) that embeds a chat assistant into the Desk and treats the AI provider as a pluggable, swappable dependency. Under the hood we define DocTypes for provider configuration, prompt templates, and per-user conversation history, with API keys stored in Frappe's encrypted Password fields rather than in code or plaintext settings. Providers (OpenAI, Anthropic, Gemini, and self-hosted Ollama) sit behind a single dispatcher class so the assistant can fail over to the next configured provider automatically when one returns an error or times out. A client script adds the chat panel to Form views and captures the current DocType and document name; a whitelisted server method (@frappe.whitelist()) injects that document's fields as grounded context into the prompt, so an answer about the invoice on screen actually references the invoice on screen.

The assistant streams responses token-by-token into the Desk using Frappe's realtime (frappe.realtime / SocketIO) layer, so long answers appear as they generate instead of after a spinner. A prompt template library — stored as a DocType and filtered by the user's Role Profile — lets you ship curated, role-specific system prompts (a different default framing for an accountant than for a sales user) instead of leaving prompt engineering to each employee. Conversation history is persisted per user with standard Frappe permissions and role-based access, so users only see their own threads and admins can audit usage. Where you need periodic work — nightly summaries, usage-cost rollups, stale-thread cleanup — we wire scheduler_events in hooks.py; where you need the assistant to react to record changes, we use doc_events hooks. Everything is reachable through the standard Frappe REST API and whitelisted methods, so the same assistant can be called from custom pages or external tools you already run.

Because this is built to order, nothing is a generic download. We start from a short scoping call, confirm exactly which providers, roles, prompt templates, and DocTypes you want in scope, then build against your ERPNext v15 or v16 site. You get the installable source for your version, a UAT pass on a staging bench with a documented rollback plan, a git repository handover, and a post-go-live support window. Typical delivery is one working week from confirmed scope, depending on how many providers and role-specific prompt sets you need wired in.