How training and retrieval work

Training turns approved source text into searchable knowledge. During a conversation, Fetchply retrieves relevant parts and gives them to the selected model with the agent's instructions.

Source states

• Processing means extraction, chunking, or indexing is still running.
• Indexed means the source can be retrieved for answers.
• Failed means the source was not made available. Open details for the reason.

Initial training can make chat available after the first source is indexed. Other sources may still be processing, so early answers can be incomplete.

What training does not do

Training does not guarantee that every answer will quote a source, resolve conflicting content, read image-only documents, or access a private site.

Instructions shape behavior; they are not a substitute for factual sources. Functions provide live external data; they are not part of the static knowledge index.

Improve answer quality

• Remove duplicate and obsolete sources.
• Give one policy a single current source of truth.
• Use Instant Answers (Q&A Pairs) for recurring questions with approved wording.
• Enable citations when visitors should inspect the supporting source.
• Test how customers actually phrase questions.

Repeated questions

Fetchply can answer an identical, standalone knowledge question faster by reusing a recent source-backed answer. The saved reply is refreshed automatically when you retrain, edit an Instant Answer, add or remove a source, or change the agent's instructions.

Replies are not reused when a question depends on the visitor, conversation history, an attachment, current commerce data, a live integration, an action, or a human handoff. These requests continue through the normal live response flow.

If a knowledge update is still processing, Fetchply waits to reuse answers until the update finishes. A failed update does not restore an older saved reply; fix the failed source and try the update again.

Troubleshooting

• A source cannot be deleted: The source remains listed and available. Check your connection and try again; do not repeatedly upload the same source as a workaround.