A RAG-powered chat assistant in the right-toolbar dock. Ask natural-language questions grounded in your entity records and text attachments. On-VM embeddings, no external API calls.

Yoker

Yoker is FastYoke's AI assistant — a RAG-powered chat that lives in the right-toolbar dock alongside Messaging. Ask natural-language questions about your tenant's entity records and their text attachments; the platform retrieves the most relevant chunks and synthesizes an answer.

Two things make Yoker different from a generic chatbot wired to your data:

Embeddings run on your VM. The platform never sends your data to OpenAI, Anthropic, or any other external embedding provider.
Conversations are ephemeral. The server doesn't persist your chat history; the client keeps it during a session and forgets it after.

How to enable Yoker

Two paths:

Enterprise / Fleet — Yoker is included. No add-on needed.
Pro / Team — Yoker requires the paid yoker add-on at $299/mo (rates as of June 2026; check the platform's pricing page for current rates).

Hobby / Solo tiers cannot use Yoker — no add-on path is offered.

Without entitlement, every Yoker endpoint returns 403 with a yoker_not_entitled error code (see API reference) and a CTA pointing at the add-on activation flow.

The dock

Yoker lives in the right-toolbar dock — the same surface Messaging uses, different tab. The dock is persistent across pages; chat stays open while you navigate.

Conversations are ephemeral. The platform doesn't store your chat history server-side — the client maintains the conversation during a session, the server forgets every turn after answering.

What Yoker knows today

Two ingestion sources are shipped:

Entity records. Every entity record's data_payload is rendered to text, chunked (sliding window 1500 chars / 200 overlap), embedded, and stored. Write-through on every entity create or update — Yoker reflects new records within seconds.
Entity file attachments — text, CSV, and Markdown only. Uploaded files with text/*, text/csv, or text/markdown MIME types are parsed UTF-8 lossy, chunked, and embedded with the source filename prefixed for context. PDF, Word, and Excel attachments are not parsed today (see deferred list).

Privacy: on-VM, no external API calls

Yoker runs all embeddings locally on the tenant's VM using ONNX runtime (tract-onnx + all-MiniLM-L6-v2, 384-dim). The platform never sends your data to external embedding providers (OpenAI, Anthropic, Voyage, or others).

The vector store is the tenant-scoped rag_chunks table — chunks live with the tenant data and are isolated by the platform's prime directive. See Tenant scoping for the full contract.

Backfill

When you enable Yoker on a tenant with existing data, the corpus starts empty until you run a backfill:

POST /api/v1/tenant/rag/backfill

The backfill walks every entity record and every text/CSV/Markdown attachment, re-embeds, and upserts. Idempotent via content_hash — running it twice is safe.

After the initial backfill, write-through keeps the corpus current for entity records (create/update) and attachments (upload/delete).

What's not in Yoker today

Documentary list of explicit non-features:

PDF, Word, and Excel attachment parsing — only text/CSV/ Markdown attachments are indexed today.
Streaming chat responses — responses arrive as a single complete message; no token-by-token streaming.
Persistent conversation history — chat history is ephemeral and client-side only.
External-embedding APIs — Yoker doesn't support sending your data to OpenAI / Anthropic / other embedding providers; everything stays on the VM.
Platform documentation as a corpus — Yoker can't answer "how do I configure forms?" from these docs today.
Workflow configuration and operational history — Yoker can't answer "what FSM transitions fired yesterday?" from job audit logs today.