Why Bedrock replies work but AI Assistant knowledge fails in WHMCS
A working Nova reply does not prove that custom Knowledge can be indexed or searched. PD AI Assistant uses a separate Bedrock embedding model to turn reference text and search questions into comparable representations.
Identify which operation failed
Open PD AI Assistant → Settings → Provider & connection and check that Bedrock is the active provider. Then open Settings → Knowledge and check the Indexing with summary.
Distinguish a failed ordinary reply from an Add to knowledge base indexing error or a reply that cannot find an expected reference. These are different checks. Keep the attempt time and redacted error rather than retrying a large document repeatedly.
Check the Bedrock key’s scope
The module’s Bedrock chat choices use Nova. Custom Knowledge uses Amazon Titan Text Embeddings V2, model amazon.titan-embed-text-v2:0, through the same saved AWS connection. AWS documents that model separately in its Titan embedding reference.
The dedicated key created by the module’s generation option is scoped to Nova models and Nova inference profiles. Its declared policy does not include Titan. If you chose that option, do not assume it supplies custom Knowledge access as well as chat.
Review whether the saved connection supports both chat and custom Knowledge, rather than treating a working Nova reply as proof of embedding access. If you manage an existing AWS key yourself, have your AWS administrator check access to the embedding model in the configured region. Do not broaden unrelated permissions or run internal repairs inside your billing system.
Check the reference itself
In Indexed sources, check the source’s Embedded with provider, Visible to audience and passage count. Sources are not rebuilt automatically when you change providers. Follow the Knowledge replacement guide to replace old references deliberately.
Use a short, harmless piece of text for a controlled indexing check once the connection is suitable. If the error concerns document extraction, try reviewed plain text instead of assuming the AWS key is at fault. After an indexing error, inspect the source list before repeating the add: an unsuccessful result is not proof that nothing was saved.
Verify retrieval after indexing
Ask a new question using a distinctive fact from the test reference and review the answer and citation. If the reference is Staff only, check it in the staff assistant. That audience setting does not make the document available to customer chat.
Successful indexing and correct audience selection help establish the setup, but every answer still needs review. For an unresolved result, retain the selected provider, source region, source title, intended audience, original indexing result and redacted error while your hosting team checks the failed stage. Keep confidential documents, credentials and private conversations out of the ticket.
See the PD AI Assistant WHMCS guide for related tasks, or explore PD AI Assistant.