RAG Architecture Planner | Leila Fielding

Runs in your browser

Use this when: documenting sources, access controls, retrieval assumptions and evaluation before implementation—not connecting or testing a live RAG system.

WHAT YOU'LL ACHIEVE

Document a human-reviewed RAG architecture using source, access, retrieval and operational assumptions you control.

THE TOOL: RAG ARCHITECTURE PLAN

Plan the evidence path.

This is a documentation planner only. It does not ingest real data, connect a vector database or model, or retrieve live information.

0 of 18 required planning decisions recorded. Still decide: Use case and intended users; Corpus and source types; Source ownership and authority; Proposed ingestion and extraction; Chunking approach and assumptions; Metadata and provenance; Embedding model assumptions; Indexing approach assumptions; Retrieval strategy; Citations and answer grounding; Access control and sensitive-data handling; Data sensitivity and exclusions; Freshness and update policy (documentation only); Evaluation set and test queries; Failure modes and fallbacks; Human review and escalation; Operational cost considerations; Acceptance gates and accountable owners.

SECTION 1 OF 5

Use case and source corpus

Define the information need and whether the candidate sources are suitable.

Use case and intended usersWhat question should the proposed retrieval system support, and for whom?Required decision.Corpus and source typesList source classes, formats, languages and known exclusions; do not enter source content.Required decision.Source ownership and authorityName the accountable owner, authoritative version and permission to use each source class.Required decision.

SECTION 2 OF 5

Ingestion and representation assumptions

Document candidate processing choices only; this planner does not ingest or index documents.

Proposed ingestion and extractionRecord how updates, formats, OCR, parsing failures and exclusions might be handled.Required decision.Chunking approach and assumptionsDescribe boundaries, overlap, size trade-offs and how headings or tables remain attached.Required decision.Metadata and provenanceConsider source ID, title, owner, effective date, version, section, classification and access labels.Required decision.

SECTION 3 OF 5

Embeddings, index and retrieval

Treat technical choices as user-supplied assumptions, not configured components.

Embedding model assumptionsRecord candidate model/version, language fit, hosting, data handling and change-control questions.Required decision.Indexing approach assumptionsDescribe the candidate index type, tenancy, access filtering, refresh and deletion behaviour.Required decision.Retrieval strategySet candidate filters, keyword/vector or hybrid approach, top-k/reranking assumptions and low-result handling.Required decision.Citations and answer groundingDefine what evidence must accompany an answer and when the system should decline to answer.Required decision.

SECTION 4 OF 5

Access, sensitivity and freshness

Keep access checks and update policies tied to source ownership.

Access control and sensitive-data handlingDescribe identity-aware filtering, least privilege, redaction, logging and deletion constraints.Required decision.Data sensitivity and exclusionsClassify likely content and document restricted categories or unresolved approvals.Required decision.Freshness and update policy (documentation only)State proposed owners, refresh triggers, target cadence, stale-content handling and removal process.Required decision.

SECTION 5 OF 5

Evaluation, failures and operational readiness

Plan evidence, human review and acceptance before any separate build decision.

Evaluation set and test queriesInclude expected evidence, answerable/unanswerable queries, permission cases and measures.Required decision.Failure modes and fallbacksConsider stale or conflicting sources, retrieval misses, parser errors, injection, leakage and service outages.Required decision.Human review and escalationName who checks answers, which decisions require review, and how corrections or incidents are handled.Required decision.Operational cost considerationsRecord expected corpus size, update frequency, query volume, embedding/re-index work, monitoring and support assumptions.Required decision.Acceptance gates and accountable ownersSet measurable quality, access, freshness, citation and operational criteria plus sign-off roles.Required decision.

HUMAN-READABLE SUMMARY

Review the proposed plan.

Decisions still needed

Use case and intended users
Corpus and source types
Source ownership and authority
Proposed ingestion and extraction
Chunking approach and assumptions
Metadata and provenance
Embedding model assumptions
Indexing approach assumptions
Retrieval strategy
Citations and answer grounding
Access control and sensitive-data handling
Data sensitivity and exclusions
Freshness and update policy (documentation only)
Evaluation set and test queries
Failure modes and fallbacks
Human review and escalation
Operational cost considerations
Acceptance gates and accountable owners

Show proposed retrieval and review flow

This optional view is a planning aid, not an executable pipeline.

Confirm use case, source authority and permitted audience.
Document extraction, chunking, metadata and index assumptions.
Apply identity-aware access filters before selecting evidence.
Ground answers in current, attributable sources; abstain when evidence is missing.
Route conflicts, sensitive cases and failures to a human owner.
Test queries and access boundaries against acceptance gates.

RAG ARCHITECTURE PLAN: DOCUMENTATION AID

USE CASE AND SOURCE CORPUS
Use case and intended users: Not decided
Corpus and source types: Not decided
Source ownership and authority: Not decided

INGESTION AND REPRESENTATION ASSUMPTIONS
Proposed ingestion and extraction: Not decided
Chunking approach and assumptions: Not decided
Metadata and provenance: Not decided

EMBEDDINGS, INDEX AND RETRIEVAL
Embedding model assumptions: Not decided
Indexing approach assumptions: Not decided
Retrieval strategy: Not decided
Citations and answer grounding: Not decided

ACCESS, SENSITIVITY AND FRESHNESS
Access control and sensitive-data handling: Not decided
Data sensitivity and exclusions: Not decided
Freshness and update policy (documentation only): Not decided

EVALUATION, FAILURES AND OPERATIONAL READINESS
Evaluation set and test queries: Not decided
Failure modes and fallbacks: Not decided
Human review and escalation: Not decided
Operational cost considerations: Not decided
Acceptance gates and accountable owners: Not decided

PROPOSED FLOW (NOT CONFIGURED)
Define a user question → verify source ownership and permissions → document ingestion and chunking assumptions → retrieve only authorised evidence → cite sources and abstain when evidence is insufficient → human review and escalation → evaluate acceptance gates.

Freshness and update policy above is documentation only. No corpus is ingested, no index or vector database is created, no embedding or model is connected, and no live retrieval occurs. Incomplete decisions: Use case and intended users; Corpus and source types; Source ownership and authority; Proposed ingestion and extraction; Chunking approach and assumptions; Metadata and provenance; Embedding model assumptions; Indexing approach assumptions; Retrieval strategy; Citations and answer grounding; Access control and sensitive-data handling; Data sensitivity and exclusions; Freshness and update policy (documentation only); Evaluation set and test queries; Failure modes and fallbacks; Human review and escalation; Operational cost considerations; Acceptance gates and accountable owners. Inputs remain in this page session only.