RAG as a Service: cited context bundles, not answers

The line Wikantik will not cross: no synthesis

A lot of "RAG platforms" quietly answer for you. You send a question, they run retrieval, feed it to a model they chose, and return prose. That is convenient right up until the answer is wrong — and now the hallucination is coming from your knowledge base, with no way to see which passage it came from or whether that passage still says what the model claimed.

Wikantik draws the line at assembly. It retrieves, ranks, de-duplicates, and cites — and then it stops. The synthesized answer is your model's job, in your agent, under your control. Every section in the bundle carries a citation to a specific page and version, so anything your agent asserts can be traced back to a source you can open. That is the whole point of grounding: not a confident paragraph, but a confident paragraph you can check.

The context bundle

The context bundle is Wikantik's core retrieval product. Ask for one over the assemble_bundle MCP tool or the GET /api/bundle?q= endpoint, and you get back a compact, ordered set of the most relevant sections across your entire wiki — not whole pages, and not a list of links to go fetch yourself. Assembly does the work that RAG pipelines normally push onto the caller:

• Section-level, whole-corpus retrieval. Candidates are drawn from a global pool of content chunks, so a highly relevant section is never dropped just because its page ranked outside the top-N. No page pre-selection to lose good passages behind.
• Hybrid scoring. Keyword (BM25) and dense-embedding similarity are fused with weighted Reciprocal Rank Fusion — the same fail-closed hybrid engine described on the hybrid retrieval page — so plain-language questions and domain jargon both land.
• De-duplication. Near-duplicate sections are collapsed, so the bundle spends its token budget on distinct information instead of the same paragraph three times.
• Version-pinned citations. Each section is stamped with its source page and the exact version it came from — the grounding your agent cites, and the anchor the self-healing loop watches.
• A coverage signal. Every bundle carries a small coverage block — how many sections and distinct pages it drew from, the top similarity, and a confidence rating. If the wiki did not really cover the question, the agent is told so up front, and can say "I don't know" or route elsewhere instead of confidently answering from three weak hits.

The result is that an agent's retrieval step is a single call that returns citable context, not a mini-project. Both surfaces return the same structure; the /api/bundle endpoint mirrors the assemble_bundle tool for pipelines that are not MCP clients.

Session briefings: grounded before the first prompt

Retrieval mid-task is reactive — the agent already has to know what to ask. A lot of the value of a team wiki is the context an agent should have before it starts: the conventions, the runbooks, the design decisions for the area it is about to work in. Wikantik packages that as a session briefing.

A coding agent — Claude Code, or any compatible assistant — requests a briefing at session start via the get_briefing MCP tool or GET /api/briefing, passing any combination of pinned pages, clusters, and the task prompt. Wikantik assembles a budgeted, de-duplicated block of the most relevant context and returns it as injection-ready Markdown (or structured JSON), which the agent drops in as opening context. The agent starts the task already knowing how your team does things — no "read these ten pages first" ritual, no blank-slate first draft that ignores your conventions. Small client shims make wiring it into an agent runtime a copy-paste.

Citations that repair themselves

Grounding is only worth something if it stays true. In Wikantik, an author (or a curator agent) marks a claim's source inline in the page body, and at save time that becomes a first-class citation edge — version-pinned and span-hashed to the exact passage it grounds.

When a cited page is later edited, Wikantik does not just invalidate everything that pointed at it. It grades staleness at the span level — did the cited passage actually change, or just something elsewhere on the page? — and surfaces what genuinely drifted through the list_stale_citations tool and a drift dashboard. Stale groundings get found and fixed instead of quietly rotting. This is the self-healing half of RAG-as-a-Service: the same version pins that make a bundle citable also make the corpus tell you when a citation has gone out of date.

Access control comes along for free

Bundles and briefings are assembled from the same retrieval candidates as the rest of the platform, so the same page-level ACLs apply. A section from a restricted page never appears in a bundle or briefing for an agent whose session cannot view it — grounding is filtered to what the caller is actually allowed to read, with no separate policy to keep in sync.

Frequently asked questions

Does Wikantik answer questions for my agent?

No — and that is deliberate. Wikantik assembles a ranked, de-duplicated context bundle of the most relevant sections with version-pinned citations, then hands it to your model. Your model does the reasoning and writes the answer. The wiki never synthesizes prose, so there is nothing to hallucinate and every claim traces back to a cited page and version.

How is a context bundle different from a plain search API?

A search API returns a list of pages and leaves ranking, de-duplication, section selection, and citation to you. A context bundle does that assembly server-side: it retrieves at the section level across the whole corpus, fuses keyword and dense-embedding scores, removes near-duplicate sections, pins each one to a page version, and returns a coverage signal telling the agent how well the knowledge base actually covered the question.

What is a session briefing?

A session briefing is a budgeted, injection-ready block of context assembled at the start of an agent session from pinned pages, clusters, and the task prompt. Coding agents such as Claude Code and compatible assistants pull a briefing via the get_briefing MCP tool or the /api/briefing endpoint and inject it as opening context, so the agent starts the task already grounded in your wiki.

How do cited sections stay accurate as pages change?

Every citation is version-pinned and span-hashed. When a cited page is edited, Wikantik grades the citation's staleness at the span level and surfaces the drift through the list_stale_citations tool and the drift dashboard, so stale groundings are found and repaired instead of silently rotting. This is the self-healing loop behind Wikantik's RAG-as-a-Service.