Platform Architecture
The runtime model behind Torra's knowledge-grounded and multimodal execution.
Runtime position
Torra is the answer and execution layer that sits on top of model providers, knowledge infrastructure, and business workflows.
Core system flow
- Files land in raw storage and are normalized into document pages, slides, and clean text.
- Normalized chunks are embedded and stored with collection and source metadata.
- Retrieval pulls candidate chunks and reranks them before answer assembly.
- Torra formats the answer, preserves citations, and reuses the evidence package for downstream generation.
Why this matters
- The retrieval layer stays replaceable.
- The answer format stays product-defined.
- The same evidence package can support ask, doc, deck, and media generation flows.
- Teams can review outputs against cited source material instead of trusting an opaque model response.
Architecture principles
- Lead with ownership of the knowledge layer.
- Keep citations visible and reviewable.
- Treat answer formatting as a product responsibility.
- Reuse one governed runtime across every team rather than splitting by model or modality.
Typical operators in the stack
- Torra ingestion for normalization
- Vector index and metadata store for retrieval
- Reranking for evidence quality
- Torra answer layer for formatting, citation, and workflow reuse