MVP Stage 0 Corpus API + MCP
Read-only, permission-aware retrieval over the existing corpus, callable from Claude, ChatGPT, or an internal copilot.
Solution and product roadmap
Build the substrate first. Productize the app later.
ReviewOS should start as a read-only retrieval layer over MediGen's existing corpus. The application comes after real usage shows which review workflows are worth standardizing.
MVP product
MVP Stage 0 is the current build target: a Corpus API and MCP server that parses the existing ~50K-document corpus, stores it in the AI semantic search engine, builds semantic and keyword indexes, enforces document-level permissions, and returns citation-first evidence objects to approved AI clients.
The boundary is explicit. MVP Stage 0 is read-only. It does not own answer synthesis, write documents, replace privilege review, or ship a new workflow UI.
Release ladder
Product value by layer
| Release | Product form | What it adds |
|---|---|---|
| MVP Stage 0 | Corpus API + MCP server | Read-only retrieval substrate over the existing corpus. |
| MVP Stage 1 | Document ingestion | Controlled corpus maintenance and refresh workflows. |
| MVP Stage 2 | Citation + cross-reference agent | Grounded synthesis, contradiction checks, and evidence packets. |
| MVP Stage 3 | ReviewOS application | Projects, queues, approvals, templates, exports, analytics, and audit. |
| X1-X3 | Extensions | Multimodal search, thinking-model expansion, and recursive retrieval optimization. |
Capabilities
What ReviewOS becomes, in order
Hybrid retrieval
Vector search plus BM25 keyword retrieval, metadata filters, exact terms, IDs, dates, clauses, and regulation references.
Citation-first evidence
Structured evidence objects return source text, document ID, title, page, chunk, score, and metadata.
MVP Stage 1 corpus maintenance
Authorized upload, batch ingestion, index refresh, duplicate detection, failed-ingestion queue, and deprecation.
MVP Stage 2 citation agent
Headless multi-pass retrieval, cross-reference checks, contradiction and gap detection, and evidence packets.
MVP Stage 3 review application
Projects, queues, approvals, templates, exports, analytics, RBAC, SSO, and full source-to-output audit.
12-week product roadmap
Six sprints, three gates
S125 pts
S229 pts
S331 pts
S428 pts
S529 pts
S628 pts
Foundations and corpus assessment
Weeks 1-2 · Foundation
Stand up repo, CI, provider abstraction, hosting/database path, corpus inventory, parsing spike, permission discovery, and eval sets.
Value: Create the factual basis for MVP Stage 0 instead of guessing.
MVP Stage 0 ingestion and indexing
Weeks 3-4 · MVP Stage 0
Parse PDF/DOCX/TXT/spreadsheets, chunk passages, extract metadata, build vector and BM25 indexes, and log parse failures.
Value: Make the existing corpus operationally queryable.
MVP Stage 0 retrieval, API, MCP
Weeks 5-6 · MVP Stage 0
Ship hybrid retrieval, metadata filters, citation-first evidence API, read-only MCP server, permissions, logs, and eval gate.
Value: Open a pilot inside existing AI tools.
MVP Stage 1 ingestion workflow
Weeks 7-8 · MVP Stage 1
Add controlled upload, batch ingestion, incremental refresh, duplicates, failed-ingestion queue, deprecation, and ingestion audit.
Value: Keep the corpus current without a full rebuild.
MVP Stage 2 citation agent core
Weeks 9-10 · MVP Stage 2
Add query decomposition, multi-pass retrieval, contradiction/gap detection, citation validation, and evidence packet output.
Value: Move from cited retrieval to grounded synthesis.
MVP Stage 2 harden and Stage 3 go/no-go
Weeks 11-12 · MVP Stage 2 + Stage 3 decision
Harden confidence, safe-fail behavior, permissions, eval dashboard, and use pilot logs to decide whether to build the Stage 3 app.
Value: Productize only the workflows proven by usage.
MVP 0-2 system architecture
How the staged solution works
Stage 0 proves trusted retrieval, Stage 1 keeps the corpus current, and Stage 2 turns cited passages into reusable evidence packets without jumping straight to a workflow app.
Trusted corpus access
Approved repositoriesParser + chunkerMetadata + permissionsVector + BM25 indexes
Existing documents become searchable, permission-aware evidence.
Citation-grade retrieval
Read-only Corpus APIMCP retrieval toolHybrid rankingCitation-first evidence objects
Reviewers get exact passages, source text, pages, scores, and filters.
Corpus maintenance
Controlled uploadBatch ingestionDuplicate + version checksFailed-ingestion queue
The search layer stays current without rebuilding the corpus by hand.
Grounded evidence packets
Query decompositionMulti-pass retrievalGap + contradiction checksCitation validation
The product moves from cited retrieval to defensible first-pass synthesis.
Reviewer
Asks a research, legal, or regulatory question in an approved AI tool.
ReviewOS
Scopes the request against permissions, metadata filters, and corpus boundaries.
Retrieval layer
Runs semantic and keyword search, ranks passages, and returns citation-first evidence objects.
Human reviewer
Validates source fit, sensitivity, and answer use; ReviewOS logs query, sources, and failures.
Stage 1 maintenance
Adds controlled ingestion, refresh, duplicates, deprecation, and ingestion audit.
Stage 2 citation agent
Builds grounded evidence packets with corroboration, contradictions, gaps, and citation checks.
Not now
Explicitly deferred
- MVP Stage 3 application build until Stage 0-2 usage proves repeated workflows.
- Autonomous legal, regulatory, or scientific judgment.
- Final privilege decisions without human review.
- Deep per-DMS integrations before API/upload/export prove value.
- Multimodal document search, fine-tuning, and retrieval optimization agents.
- Slack/Teams, external filing, and autonomous stakeholder communication.