Upload policies, rate cards, decks — even recorded training calls. Zoice chunks and embeds them for hybrid RAG retrieval, tags every chunk by language, and has agents append a "Sources:" footer to grounded answers.
loan-products-kb
Attached to 3 agents
Home-Loan-Policy.pdf
482 chunks · hi / en
Branch-Training-Call.mp3
Transcribed · 96 chunks
Rate-Card-Q2.xlsx
38 chunks · en
616
Total chunks
100%
Embedding coverage
412
Avg tokens / chunk
Rechunk recommended
2 files changed since last index
Chunk health updates automatically as your documents change.
Everything between upload and a cited answer is visible — chunk counts, retrieval rankings, and grounding checks included.
Drop in .pdf, .txt, .doc/.docx, .ppt/.pptx, .xls/.xlsx, .csv, .json, or .md files — or paste raw text straight into the editor. Audio and video (.mp3, .wav, .ogg, .m4a, .mp4, .webm) are transcribed first, so a recorded training call becomes searchable knowledge.
Zoice splits documents into chunks, embeds them for hybrid RAG retrieval, and tags each chunk by language. Chunk health stats — total chunks, embedding coverage, average chunk size — stay visible, with an automatic "rechunk recommended" hint when the index drifts.
Before going live, run a real query in the retrieval test bench and inspect the exact top-K chunks the agent would receive — same content, same ranking. No guessing whether the right passage will surface on a live call.
Attach knowledge bases to any agent in the editor — voice, WhatsApp, or web chat. Agents append a "Sources:" citation footer to grounded answers, and LLM-as-judge hallucination detection can be toggled per agent under Trust & Compliance.
Not a black-box vector store — every stage from ingestion to citation is inspectable.
Upload .pdf, .txt, .doc/.docx, .ppt/.pptx, .xls/.xlsx, .csv, .json, and .md files — or skip files entirely and paste raw text straight into a knowledge base.
Drop in .mp3, .wav, .ogg, .m4a, .mp4, or .webm files. Zoice transcribes them into searchable chunks, so yesterday's recorded call can answer tomorrow's customer.
Semantic embeddings and keyword matching work together, so exact terms like policy codes and product SKUs surface alongside loosely-phrased questions.
Each chunk carries a language tag, so a Hindi question retrieves Hindi answers instead of a translation-mangled English passage — built for agents speaking 10+ Indian languages.
Total chunks, embedding coverage, and average chunk size at a glance — plus an automatic "rechunk recommended" hint when documents change and the index drifts.
An LLM-as-judge checks whether each answer is grounded in retrieved chunks. Ungrounded answers land in a triage dashboard with low/medium/high severity and an acknowledge/dismiss workflow.
Walk the exact pipeline your documents travel — and see why agents stop guessing.
Drop files — or paste raw text
Audio and video are transcribed before chunking.
Home-Loan-Policy.pdf
482 chunks · hi / en
Branch-Training-Call.mp3
Transcribed · 96 chunks
Rate-Card-Q2.xlsx
38 chunks · en
Rechunk recommended — chunk health (total chunks, embedding coverage, average chunk size) is tracked per knowledge base.
The same documents ground your phone calls, chats, and dashboards.
Hear the knowledge base at work — AI agents answering on real phone lines.
ExploreSee which answers land — sentiment, objectives, and Ask Anything over your calls.
ExploreHallucination triage lives beside consent capture, retention, and access controls.
ExploreNew to Zoice? Start from the home page for the full platform tour.
Upload your documents, test retrieval live, and ship agents that cite their sources on every grounded answer.