Typical answer-oriented RAG
- Question
- Retrieve relevant chunks
- LLM answer
RAG AND EVIDENCE-BOUND OPERATIONS
Orbivell uses retrieval where it works best—then adds exact source locations, cross-source checks, repeatable numeric and date checks, revision-impact tracking, user approval, editable deliverables, and review history.
FAIR COMPARISON
| Area | Typical RAG | Orbivell |
|---|---|---|
| Primary outcome | Answer linked to source passages | Verified, approved work result and editable file |
| Source detail | Relevant passages and document citations | Document version, page, clause, table, cell, and field locations |
| Structured-data reconciliation | Optional and implementation-dependent | Cross-checks across documents, Excel, CSV, and approved read-only data |
| Repeatable numeric/date checks | Model explanation or separate logic | Reproducible rules recalculate values |
| Conflict handling | The model may synthesize | Conflicting, ambiguous, and unverifiable evidence remains explicit |
| Revision impact | Usually retrieve or generate again | Changed-source impact remains visible in obligations, reports, and slides |
| Execution control | Implementation-dependent | Preview → approval → bounded execution |
| Editable outputs | Answers, summaries, and drafts | Editable PPTX, DOCX, and PDF |
| Source protection | Product-dependent | No source write, move, delete, rename, or overwrite |
| Reviewability | Conversation and citation records may be available | Source and approval records with tamper checks |
RAG implementations vary. Advanced Graph RAG and agentic RAG may include more capabilities; this page compares a typical answer-oriented pattern.
Simple document Q&A, fast summaries, FAQ discovery, and draft ideas.
Contract expiry and notices, document-to-spreadsheet checks, formal reports, regulation impact, editable external deliverables, and audit evidence.
WORKED EXAMPLE
FAQ
Yes. RAG is part of the retrieval layer.
No. Implementations vary; this comparison uses a typical answer-oriented pattern.
Numbers, dates, permissions, paths, and approvals require reproducible rules.
No. Local processing is the default; managed AI receives only explicitly approved evidence.
Yes, according to the licensed plan and verified renderer support.
Optional. We record only page category, language, edition, channel, campaign code, version, time, and a server-hashed pseudonym. We never record document names, paths, questions, answers, inquiry text, email, or payment instruments. Declining does not limit any product feature.