Isolation
Website-scoped knowledge
Pages, prepared knowledge, questions, and answers remain tied to the active website instead of mixing unrelated customer content.
Technology
RAG + voice readinessGrounded answers first. Natural voice delivery next.
Guftara separates website preparation, grounded answering, and speech delivery so each part can be evaluated, secured, and scaled for a real pilot.
System flow
A simple path from source content to conversation.
The current MVP completes the website-to-grounded-answer path. Pilot infrastructure will add and qualify multilingual speech delivery.
- 01
Scan
Read permitted public website pages and preserve a clear record of each scan attempt.
- 02
Prepare
Turn successful pages into website-scoped knowledge that can support later questions.
- 03
Ground
Find relevant source context and compose an answer within the active website boundary.
- 04
Deliver
Return a reviewable text answer today and qualify multilingual speech delivery during pilots.
Technical foundations
Built around boundaries that make answers reviewable.
Quality
Grounded answer review
The Preview Assistant can surface supporting page titles and URLs so a user can inspect where an answer came from.
Language
Multilingual by design
The architecture is designed for multilingual retrieval and speech. Exact language and voice support is qualified and published for each pilot.
Scale
Asynchronous preparation
Long-running website preparation is separated from interactive questions so each workload can scale independently.
Why cloud credits matter
Pilot workloads need more than a web server.
Credits extend runway while Guftara measures model quality, speech latency, and real usage before committing to long-term capacity.
Application compute
Runs the product API, account workflows, website coordination, and background preparation jobs.
Durable data and retrieval
Stores website records, Indexed Pages, scan state, and vector-capable representations used to ground answers.
Model inference
Supports embeddings, answer generation, evaluation, and controlled experiments with managed or open models.
Speech acceleration
GPU capacity is needed to evaluate low-latency multilingual text-to-speech models for pilot conversations.
Evaluate the product against a real website.
A focused pilot helps us measure grounding quality, language needs, response latency, and the infrastructure required for production.