RAG-as-a-Service Roadmap Hub

This hub indexes the long-term technology roadmap that turns Wikantik from "a wiki with excellent agent surfaces" into a credible competitor in the RAG-as-a-service market. It was produced from a 2026-07-10 market gap analysis comparing Wikantik against commercial and open-source offerings: Glean, Onyx, Amazon Bedrock Knowledge Bases, Azure AI Search, Vectara, Ragie, LlamaCloud/LlamaIndex, Pinecone, and Unstructured.

Each linked page is a one-page capability brief designed to bootstrap a deep-dive investigation by an agent or a team: gap statement, market grounding with sources, current Wikantik state with code and design-doc entry points, a proposed direction, Phase-1 acceptance criteria with a first-session scope, a verify-before-designing claims list, non-goals, and tagged open questions. The expected next step for any brief is the project's standard process: brainstorm → design doc → phased implementation.

Strategic thesis: the curated, agent-native knowledge base

Wikantik should not chase Glean's 100-plus-connector arms race or Bedrock's managed-everything on their terms. The defensible position, supported by the market research, is the self-hosted, agent-native, curated knowledge base: content that is verifiable (version-pinned citations, span-hashed staleness tracking, page verification workflow), machine-legible (RDF/SPARQL ontology, frontmatter, structural spine), and served through the protocol agents actually use (MCP). Practitioner reports describe a migration away from server-side prompt-stuffing pipelines toward MCP-mediated retrieval where the agent drives its own search — Wikantik is already built for that world.

Three existing moats to defend and market, not rebuild:

  1. Citation staleness and drift tracking — no surveyed commercial platform has an equivalent of version-pinned, span-hashed citations with graded staleness. Data freshness is named the key open challenge for agentic knowledge bases.
  2. The no-synthesis context bundle (ADR-0001) — Wikantik assembles cited evidence and never fabricates answers, which structurally exempts it from the hallucination-evaluation problem competitors must tool around.
  3. MCP-first agent surfaces — two production MCP servers plus an OpenAPI tool server, ahead of every open-source wiki and most commercial RAG products.

The Tier-1 gaps below are what make the RAG-as-a-service claim credible; the rest are parity checkboxes buyers will checklist.

Architectural guidance: one target state, reached incrementally

RoadmapTargetArchitecture is the architectural spine binding the briefs together. It defines the end state as an incremental evolution of the current modular monolith — new satellite modules (wikantik-connectors, wikantik-authz-sync), two retrieval pipeline seams (QueryPlanner pre-retrieval, RerankChain post-fusion), authorization pushdown at the single existing enforcement chokepoint, PostgreSQL-first for all new state, and the deployment boundary as the tenant boundary. It also fixes the eight invariants (no synthesis, fail-closed retrieval, async indexing, one authz chokepoint, PostgreSQL-first, no wikantik-main growth, measure-first, complete Apache 2.0 core) that any capability design must respect, and a capability-to-seam map showing where each brief's implementation attaches. Read it before starting a design doc for any brief.

Decision authority: who resolves an open question

Every open question in the capability briefs carries one of three tags. Agents executing a brief MUST respect them:

Two standing owner gates beyond the tags: the multi-tenancy posture ADR (RoadmapMultiTenancyPosture) may be drafted by an agent but is only ratified by the owner; and RoadmapPackagingAndPricing is owner-gated end-to-end — agents prepare decision inputs there, never binding designs.

Capability briefs by priority tier

TierCapabilityHorizonWhy this tier
1RoadmapConnectorFrameworkNow (0–3 mo design, first connectors)Widest gap; every source names connector breadth the primary differentiator
1RoadmapExternalAclInheritanceNext (3–9 mo)Permission-aware retrieval from connected sources is the enterprise gate; co-evolves with connectors
1RoadmapMultiTenancyPostureLater (9–18 mo), decision nowDefines the product shape: silo-per-tenant fleet vs shared-store multi-tenancy
2RoadmapComposableRerankingNow (cheap, measurable)Deterministic MMR/metadata-boost stages are low-cost parity with Vectara chain reranking
2RoadmapAgenticRetrievalNextQuery decomposition + sufficiency evaluation is now table stakes in Bedrock and Azure demos
2RoadmapMultimodalIngestionNextDiagrams and scanned documents are invisible to the current text-only pipeline
2RoadmapRetrievalEvaluationObservabilityNow (phase 1)Internal harness exists; productizing it is differentiation, not new science
3RoadmapCompliancePostureNext (whitepaper) → Later (SOC 2)Procurement table stakes for the commercial path
3RoadmapPackagingAndPricingLaterOpen-core packaging follows, not precedes, the Tier-1 capabilities; owner-gated end-to-end
3RoadmapClientSdkDeveloperExperienceNow (quick win)The LangChain/LlamaIndex retriever adapter is the genuine quick win; the full DX program is multi-session

Dependency graph

Execution ordering constraints beyond the tier table — start nothing whose upstream box is unresolved:

ConnectorFramework (schema carries source ACL refs from v1)
        │
        ▼
ExternalAclInheritance (enforcement over those refs)
        │
        ▼
CompliancePosture (retrieval audit + GDPR delete propagation ride connector tombstones)

MultiTenancyPosture: cost-floor measurement [MEASURE] ──▶ PackagingAndPricing (unit economics)
MultiTenancyPosture: posture ADR [OWNER] ──▶ tenant-in-a-box + fleet control plane work

RetrievalEvaluationObservability phase 1 (scheduled harness) ──gates──▶
        ComposableReranking, AgenticRetrieval, MultimodalIngestion (each accepted only on harness evidence)

ComposableReranking (RerankChain SPI) ──▶ AgenticRetrieval (decomposed results feed the chain)

TargetArchitecture API-stability contract ──▶ ClientSdkDeveloperExperience (SDK freezes the wire format)

MultimodalIngestion — independent once its diagram eval-set extension exists

Sequencing rationale

Sourcing note for downstream agents

Market claims in these briefs come from a 2026-07-10 research sweep over 18 sources (vendor documentation from Glean, Onyx, AWS, Azure, Vectara; multi-tenant RAG reference architectures; evaluation-tooling surveys). Claims were extracted from primary sources but not independently cross-verified — treat specific counts and prices as approximate. Each brief now carries a "Verify before designing on these" list isolating the claims whose falsity would change the design; verify those, and only those, before the design doc.

See Also