Platform · AI training

Train the AI on your own words.

Knowledge base, FAQs, prompts, catalog embeddings. Everything that shapes how the AI answers for your business, continuously tuned by MessageMind's prompt engineering team.

  • Upload documents and FAQ pairs
  • Custom system prompt, tone and persona
  • Catalog embeddings for ecommerce accounts
Try MessageMind free Jump to reference

Watch the walkthrough

AI training in a few minutes

What's inside

The headlines before you dive in.

  • Upload documents (PDFs, text files, policy docs) into the knowledge base so the AI can quote verbatim from your own source material.

  • Add FAQ pairs through the dashboard; each question/answer pair is written into the assembled training document and preserved byte-identically on every save.

  • Edit the custom system prompt that frames the agent (business context, boundaries, what to refuse, what to escalate).

  • Set tone and persona (voice, formality, verbosity) on top of the base prompt so answers sound like your brand.

  • Multilingual by default, since the base model answers the shopper in the language they wrote in without a per-language model swap.

  • Rebuild catalog embeddings on every detected change for eCommerce models, so new, edited or deleted products are reflected in retrieval on the next sync pass.

Documentation Platform fundamentals AI training

Platform integration

AI training

  • CategoryPlatform

Overview

MessageMind trains your AI agent on top of a shared base language model that is tuned per account against the model category you pick on setup (eCommerce, Services or Hotels). Training is driven by your knowledge base: uploaded documents, FAQ pairs, catalog data from connected integrations and the system prompt that pins the agent's tone, persona and escalation rules. Everything you add is parsed into a single assembled training document the AI reads at every turn, with product catalogs rebuilt as embeddings for retrieval. The underlying prompts and routing are tuned continuously by MessageMind's prompt engineering team as new accounts, categories and edge cases surface, so the agent gets better without you re-training it.

What MessageMind can do with it

  • Upload documents (PDFs, text files, policy docs) into the knowledge base so the AI can quote verbatim from your own source material.
  • Add FAQ pairs through the dashboard; each question/answer pair is written into the assembled training document and preserved byte-identically on every save.
  • Edit the custom system prompt that frames the agent (business context, boundaries, what to refuse, what to escalate).
  • Set tone and persona (voice, formality, verbosity) on top of the base prompt so answers sound like your brand.
  • Multilingual by default, since the base model answers the shopper in the language they wrote in without a per-language model swap.
  • Rebuild catalog embeddings on every detected change for eCommerce models, so new, edited or deleted products are reflected in retrieval on the next sync pass.

Requirements

  • A MessageMind account with an AI model configured.
  • A model category picked on setup (eCommerce for Physical Goods, Services, or Hotels); the category gates which training surfaces and integrations are available.

Available data and actions

Reads

  • Uploaded documents in the knowledge base (files and text editor entries).
  • The FAQ library (every question/answer pair saved on the account).
  • The product catalog from any connected eCommerce integration, including category, price, currency, SKU, attributes and pricing combinations.
  • The host's working hours and escalation rules, used to decide when the AI should hand off instead of answering.

Writes

  • Catalog embeddings rebuilt on every pass that detects a change.
  • The assembled, cached knowledge-base document the AI reads at inference time.
  • Trained-vocabulary tokens derived from your uploaded content and FAQ entries.

AI agent use cases

  • An eCommerce merchant uploads a returns policy PDF; the AI answers return questions with wording lifted straight from the merchant's own document.
  • A services business pastes 40 FAQ pairs about pricing, service area and booking lead times; the AI quotes those exact answers on WhatsApp, Instagram and the website chat.
  • A hotel edits the system prompt to always upsell breakfast with any room enquiry; every room answer now includes the breakfast line without a code change.
  • A merchant pushes a product edit through Cin7 Core or Shopify; the next sync rebuilds the catalog embeddings so the AI quotes the new price without a manual retrain.
  • A tenant switches tone from 'professional' to 'casual' before a holiday campaign; the same knowledge base now answers in a friendlier voice across every channel.

Configuration

  • Temperature, tone and language controls shape how the agent speaks on top of the shared base prompt.
  • Auto-handoff thresholds (confidence, keyword, working-hours) decide when the AI stops answering and routes the conversation to a human.
  • Fallback behaviour when the AI cannot answer: ask a clarifying question, offer to connect a human, or stay silent for a human to pick up, depending on the channel.

Example workflows

Add a new FAQ and test it in the sandbox

  1. Open the dashboard and go to the AI model's knowledge base section.
  2. Add a new FAQ pair (the shopper's question, the exact answer you want the AI to give).
  3. Save. The entry is written into the assembled training document and the cached knowledge base is refreshed.
  4. Open the sandbox / test chat and ask the question in the shopper's own words to confirm the AI pulls the new answer.
  5. If the answer is off, tighten the FAQ wording or add a second phrasing of the same question and re-test.

Limitations

  • The base model itself cannot be fine-tuned from the dashboard; training only shapes the retrieval layer (knowledge base, FAQ, catalog, prompts and tone), not the underlying weights.
  • The assembled training document has size caps; very large knowledge bases need to be trimmed or split rather than pasted wholesale.
  • Refresh cadence depends on the source: FAQ and prompt edits are picked up on save, but catalog embeddings only rebuild on the integration's own sync pass (daily for most eCommerce integrations).

Troubleshooting

A new FAQ is not surfacing in the AI's answers.

Confirm the save succeeded and ask the question in several phrasings in the sandbox. If the answer still does not match, tighten the FAQ wording so the question is distinctive, or add a second phrasing of the same question to improve retrieval.

An edited product price is not quoted by the AI.

Catalog embeddings rebuild on the integration's sync cadence (typically daily). Wait for the next pass or trigger a manual sync from the integration card if the integration supports it.

The AI answers in the wrong tone after a prompt edit.

Open the system prompt and confirm the tone/persona lines were saved on the right model. Tone settings are per-model, so an edit on a different AI model will not affect the one serving your channel.