Skip to main content

Understand Wirevox RAG

Wirevox uses semantic retrieval: it compares the meaning of a query with indexed pieces of source text rather than requiring exact keyword matches.

Ingestion path​

File, public page, text, or approved Vox image analysis
                         ↓
                 Extract readable text
                         ↓
              Clean and organize the text
                         ↓
           Split into overlapping text chunks
                         ↓
       Create an embedding for every text chunk
                         ↓
           Store chunks and vectors by source

Wirevox currently uses gpt-4o-mini to organize longer extracted text before chunking and OpenAI text-embedding-3-small to create 1,536-dimension vectors. It stores those vectors with the source chunks and uses cosine similarity for retrieval.

This indexing stack is independent of the conversational model. An agent can answer with a selected OpenAI, Grok, Gemini, or supported realtime model while Wirevox continues using the current OpenAI preprocessing and embedding services for knowledge retrieval.

OpenAI describes embeddings as numerical representations used to rank text by relatedness, including semantic search. See the official OpenAI embeddings guide.

Retrieval path​

Person asks a question
          ↓
Wirevox embeds the search query
          ↓
Vector search ranks chunks in the selected knowledge scope
          ↓
Results below the relevance threshold are excluded
          ↓
Up to three relevant chunks are supplied to the agent
          ↓
The selected conversational model writes or speaks the answer

The embedding service finds related text; it does not decide whether a business claim is safe to make. The system prompt and knowledge-base instructions define that policy.

Modular and Realtime voice retrieval​

For a Modular voice agent with an agent-level knowledge base, Wirevox automatically searches on meaningful caller turns, skipping very short conversational replies. The search runs with a short latency limit, and recent retrieval context is retained for follow-up turns. The model also receives a knowledge-search tool for explicit lookups.

Realtime voice agents use a knowledge-search tool during the session. Retrieval results are returned to the realtime model, which then produces the spoken answer.

The result should be equivalent at the policy level—answer from retrieved approved material or use the defined fallback—but latency and model tool-use behavior can differ. Test both after changing architecture.

Website Chat retrieval​

For the current Website Chat runtime, attach a reusable Knowledge function to expose retrieval as a callable tool. Select the required base and optionally its allowed sources in that function. Do not rely only on the Chat Configuration knowledge selection when retrieval is required; that selection supplies the base's instructions, while the Knowledge function supplies the searchable capability.

Agent-level and named retrieval​

Wirevox supports two retrieval scopes:

  • Agent-level knowledge base — one base selected in the agent configuration; voice runtimes can search it broadly.
  • Knowledge function — a named function that searches one base and optionally only selected source names; supported for voice and chat.

Use a named Knowledge function when the agent needs a clear callable purpose, a restricted source set, multiple separately named knowledge domains, or Website Chat retrieval.

Knowledge retrieval searches the indexed content already stored in the workspace. It does not browse the current public web. URL sources remain as current as their latest successful crawl. Re-crawl them when the underlying page changes.

Vox web research is also separate. Research results become persistent knowledge only after an approved create or add-sources action succeeds.

Next step​

Continue with Create and manage a knowledge base.