Two things stand between an AI voice agent and a correct answer: it has to hear the caller accurately, and it has to find the right passage in your documents before it speaks. If either fails, the agent either mishears the question or invents the answer — and both failures land in the same place, as a bad customer experience you own. This post walks through the two halves as they ship in Devotel Orbit: the STT Playground, where you compare speech-to-text providers on your own audio before you commit, and the knowledge-base pipeline, where your documents become the ground an agent answers from. Along the way it covers where each surface lives in the dashboard, who owns the credentials and data retention, and how to run the same evaluation against ElevenLabs, Vapi, or Retell.
What the STT Playground does
Voice → STT Playground exists so you never have to wire a speech-to-text provider into a live agent to find out how it handles your audio. Record a clip from your microphone in the browser, or upload one (MP3, WAV, OGG, AAC, or M4A), and every speech-to-text provider your organization has access to transcribes it at once. The results come back side by side: each provider's transcript, confidence score, and response time on the same clip, so a comparison is fair by construction. A provider that isn't configured on your account is labeled as not yet available instead of silently dropping out, so you always know what the comparison actually covered.
Two properties make the playground useful as a pre-purchase instrument rather than a demo:
- It runs on your own credential as soon as you have one. Register and activate your Deepgram API key under Settings → Voice → Speech-to-text and every playground run resolves that key — the same routing the rest of call transcription follows — falling back to the platform credential only when no active key is registered. That means the transcript and latency you measure are what your production billing would see, not a platform sample.
- A failing run tells you where to fix it. When a comparison fails because a provider's credential is missing or invalid, the failing provider card carries a "Manage the STT API key" link straight to the settings page where the key is registered, activated, and rotated — no hunting through settings.
Nothing you record or upload is saved. Each run is a scratch comparison: the clip is discarded when you close or replace it, which makes the playground safe to use with a real customer call excerpt during evaluation.
Knowledge bases as a buying decision
Once hearing works, the question becomes what the agent may say. A knowledge base is the answer to that — but not all knowledge-base features are equivalent, and for a buyer comparing agent platforms the pipeline is the product. In Orbit (Agents → Knowledge Base in the dashboard, and the AI section alongside it), the pipeline runs in four stages, and each one is a decision you control:
- Ingest and chunking. You upload documents — converted from your help center, policy PDFs, pricing sheets, or connected sources — and each document is split into passages. Chunking is the first quality gate: chunks too large drown the relevant sentence in noise, chunks too small lose the context that makes a policy answerable. The Knowledge Bases API exposes the full document lifecycle — upload, moderation, versioning — so ingestion can be automated from wherever your source content lives, not just uploaded by hand in the dashboard.
- Embeddings and indexing. Each passage is embedded and indexed for semantic search, so retrieval matches on meaning rather than keyword overlap. A caller asking "how long do I have to send it back" finds your returns policy even when the document says "return window."
- Grounding and moderation. Retrieval is scoped: an agent only draws from the knowledge-base collections you attach to it, no scope means no retrieval. Documents publish through a moderation step, and every version is kept with a diff, so what the agent could have said at any point is reconstructable — and a bad edit rolls back instead of propagating.
- Agents close the loop. When the agent answers, retrieval marks which sources it used, and the grounding audit (Agents → RAG Grounding Citations) reports the rate at which answers carry source markers, listing ungrounded turns for triage. An ungrounded cluster on one topic means the knowledge base is missing the entry — the fix is content, not prompt tuning. The sibling post on RAG grounding with citations walks that audit and remediation loop in detail.
Where these surfaces live in the dashboard
Both halves run where your team already works, so evaluation and production use the same screens:
- Voice → STT Playground — record-or-upload comparisons across speech-to-text providers, with per-provider transcript, confidence, and latency.
- Settings → Voice → Speech-to-text — the bring-your-own-key credential registration the playground and live transcription share; activate the key here and both start using it.
- Agents → Knowledge Base — document lifecycle: upload, moderate, publish, version, and the gap views that show what callers asked and no document answered.
- AI section — the AI workspace alongside Agents, where knowledge and model behavior are configured and observed.
- Agents → RAG Grounding Citations — the audit page: grounding rate per agent, ungrounded turns, and the trend that tells you whether a knowledge-base fix moved the number.
The Knowledge Bases API mirrors the dashboard surface, so everything above — upload through moderation through search through version diff — is scriptable for teams who treat documentation as code.
Tenant-owned controls: credentials and retention
Two ownership questions separate platforms you can put in front of customers from platforms you cannot, and both are answered in the tenant's favor here.
Credential ownership. Speech-to-text runs on your own registered provider key the moment one is active — your provider relationship, your pricing, your capacity — with the platform credential as fallback only when you have not brought one. The playground proves the routing before an agent ever picks up a call.
Data ownership and retention. Retention in Orbit is tenant-configured, not platform-imposed: you set the per-category retention periods — recordings, transcripts, message bodies, audit logs, knowledge and AI data — under your compliance and settings surfaces, and the platform enforces them, including legal holds and the redaction controls that strip message or transcript bodies after a window you choose while keeping the metadata. The playground side follows the same principle by construction: comparison clips are never stored at all. For a buyer, the question to bring to any vendor on this list is "who decides how long my documents, transcripts, and recordings live, and can I prove the deletion happened?" In Orbit the answer is: you decide, per category, and the audit surfaces show it.
How to evaluate this against ElevenLabs, Vapi, or Retell
Buyers comparing Devotel Orbit against the voice-agent specialists usually discover the two halves of this post are split across vendors: one tool for speech-to-text accuracy, another for retrieval, a third for the telephony underneath, with glue code holding them together. When you evaluate, run the same clip and the same document set through each platform and ask four questions:
- Can I test hearing before I ship? An STT comparison across providers, on your audio, before a live agent touches a caller — in Orbit that is the playground, not a pilot project.
- Does retrieval scope bind the agent? Attach a collection and the agent answers from it; detach it and it cannot. Answers should trace to passages, and ungrounded turns should be enumerable, not anecdotal.
- Whose credentials run the pipeline? Bring-your-own-key on the speech-to-text side with routing you can verify in one run; the playground failing forward into the right settings page is what operable ownership looks like.
- Who owns retention and deletion? Tenant-set retention per category, with evidence, beats a vendor PDF that says "we take data protection seriously."
Orbit's position is that these are one platform question, not four vendor questions — voice, speech-to-text, knowledge, and agents in one surface, where the same clip you compared in the playground and the same documents you moderated in the knowledge base are the ones the agent hears and cites in production.
Frequently asked questions
What does the STT Playground in Devotel Orbit do?
It transcribes one audio clip — recorded from your microphone or uploaded — with every speech-to-text provider available to your organization, and shows the transcript, confidence, and latency side by side. Runs use your own registered provider key once one is active, nothing is saved, and a provider without a valid credential is labeled with a link to the settings page that manages that key.
How does a knowledge base stop an AI agent from making things up?
It bounds what the agent can draw on. Documents you upload are chunked, embedded, and retrieved semantically; the agent is scoped to specific collections; and answers carry source markers that the grounding audit counts per turn. Ungrounded turns are listed so you can add the missing content — retrieval scope plus citation evidence, not a prompt that asks the model to behave.
Can I run the STT Playground on my own provider API key?
Yes. Register and activate your Deepgram key under Settings → Voice → Speech-to-text and every playground comparison resolves your key, matching the routing live transcription uses. It falls back to the platform credential only when no active key is registered.
Where do knowledge bases live in the Orbit dashboard?
Under Agents → Knowledge Base in the dashboard, with the AI section alongside it, and through the Knowledge Bases API for automation — upload, moderation, semantic search, and version diffs are all scriptable.
Who controls how long voice and knowledge data is retained?
You do. Retention periods are tenant-configured per data category — recordings, transcripts, message bodies, audit logs — with legal holds and redaction controls on top, and playground comparison clips are discarded immediately by design.