PrismVox knowledge and MCP

RAG lets the receptionist answer from reference documents. MCP lets it act through external tools. Start with a small collection and one harmless tool, then validate their use in text and on a real call. A model must support structured tool calls to use either capability.

1. Create a knowledge collection

• Open Knowledge Base → + New collection. Enter a name and optional description, then Create. Group documents by subject so they remain maintainable.
• Upload a TXT, Markdown or text-bearing PDF file in that collection. A scanned PDF with no extractable text needs OCR before upload.
• Check the chunk count and whether embeddings were calculated. Upload splits the document and attempts indexing automatically. Enabled collections supply the knowledge search.
• In Configuration → Agent enable knowledge search and save the form. Test a question from the document and a question it cannot answer.

How the agent searches

The inbound standalone agent receives search_knowledge_base when retrieval is enabled. It chooses when to call it and answers from returned passages; documents are not automatically inserted into every conversation turn. RAG_TOP_K limits the number of relevant chunks returned. Outbound missions do not receive this knowledge-search tool; put the necessary facts in their mission.

Maintain and re-index documents

Uploading the same source filename into the same collection replaces that source’s existing chunks. You can inspect/delete chunks and enable or disable collections. Re-index embeddings retries the embedding calculation, for example after the embedding service becomes available. If embedding fails during upload, text may still be stored without vectors; uploaded does not prove semantic indexing succeeded.

Standalone PrismVox uses its own local embedding service and a 384-dimensional pgvector schema. Do not replace the embedding model with an incompatible dimension. Keep document originals and database backups separately; deleting a collection removes its stored content.

2. Connect a tool server

• Open MCP → + Add an MCP server. Set a name, Type (Streamable HTTP or Legacy SSE) and the URL reachable from the PrismVox containers. Stdio is not supported.
• If required, enter its Bearer token in the secure form, then Add. This connection does not implement Prism’s browser OAuth authorization flow.
• Click Probe to discover tools. Open the server’s tool list and enable the tools the receptionist needs. Check descriptions and effects before enabling write actions.
• Keep the server enabled. In Configuration → Agent enable MCP tools and save. Test a harmless lookup in Test the agent, then repeat it over the phone.

Refresh tools and apply changes

After updating a server, Probe refreshes discovered names/descriptions and removes stale entries; inspect per-tool switches again. Disabling the server retains its configuration but withdraws its tools. Changes affect subsequent calls or text turns; an active call keeps its current MCP tool configuration. An unknown telephone caller can influence tool requests, so expose only suitable operations.

3. Verify real outcomes

Test the agent uses the standalone persona, directory and enabled RAG/MCP with the standard model, but no telephone transfer/message/hang-up tools. Admin chat uses the on-call model for administration and is a different test. A text answer does not validate SIP, transcription or audio playback. Check the tool trace and the external service’s actual result.

When connected to Prism

For inbound calls, knowledge searches use Prism’s selected public voice collection for unknown callers; recognized callers use their permitted Prism context. Local Knowledge Base collections are not merged into that public collection. Unknown callers still use the PrismVox conversation model and may use its enabled MCP tools. Recognized callers use Prism’s restricted voice tools, without inheriting the web MCP catalogue.