Skip to content
GRPNR.

R6 — Knowledge bases, Q&A products, glossary practice (2025-2026)

Research date: 2026-07-18. Purpose: does any existing product close the loop question -> human answer -> durable artifact -> next asker self-serves, for an internal 2-human + 64-agent setup.

1. Stack Overflow for Teams → “Stack Internal” (VERIFIED STATUS)

  • Rebrand, not sunset. Stack Overflow announced a full company rebrand on 2025-05-08 (stackoverflow.blog/2025/05/08/a-new-look-for-whats-next/). Quote: “with that growth, our brand began to stretch—and at times, to lose fidelity and clarity.” Feedback window ran to 2025-05-22. No product discontinuation announced in that post; it explicitly lists Stack Internal, Stack Data Licensing, Stack Ads, Teams as continuing lines. Source: https://stackoverflow.blog/2025/05/08/a-new-look-for-whats-next/ (2025-05-08)
  • Product itself renamed: “Stack Overflow for Teams” → “Stack Internal” (knowledge engine), showcased around Microsoft Ignite late 2025. Canonical page now https://stackoverflow.co/internal/. Old brand still referenced in some marketing/aggregator sites as of 2026 (inconsistent update lag across 3rd-party review sites: G2/TrustRadius/Capterra pages still say “Stack Overflow for Teams” pricing as of 2025-2026 listings). Source: https://stackoverflow.co/internal/ ; https://www.devopsdigest.com/stack-overflow-introduces-stack-internal (2025)
  • Public Q&A side of Stack Overflow (NOT Teams/Internal) is collapsing — strong negative evidence AI is eating the public commons this product depended on for network effects:
  • Negative evidence specific to the Teams/Internal enterprise product (the relevant one for internal KB use): Gartner Peer Insights reviews (2025-2026, page last updated Feb 2025) surface a recurring complaint that directly hits our “durable artifact” question: “Some answers become outdated over time… there isn’t always a clear signal that a solution is no longer valid” and “a drawback of Stack Overflow for Teams is its Q&A focus, which can result in less structured documentation.” Source: https://www.gartner.com/reviews/product/stack-overflow-for-teams (accessed 2026-07-18)
  • Verdict: Stack Internal is alive, rebranded, not sunset, no pricing-page evidence of an actual price change beyond the rename. But it has NO built-in staleness/decay signal — the exact opposite of what “durable artifact” needs. It captures the Q&A but does not maintain it.

2. Glean

  • Series F: $150M at $7.2B valuation, June 2025 (9 months after Series E at $4.6B). Wellington Management led. Source: https://www.glean.com/press/glean-raises-150m-series-f-at-7-2b-valuation-to-accelerate-enterprise-ai-agent-innovation-globally (2025)
  • ARR trajectory: ~$100M (~15 months prior) → $208M end of 2025 → $300M by May 2026 (89% YoY growth). Valuation multiple ~24x ARR — modest vs Sierra (~100x) or Harvey (~58x), i.e. market pricing Glean as infra, not magic. Source: https://www.cnbc.com/2026/05/19/glean-cnbc-disruptor-50-ranking.html (2026-05-19); https://sacra.com/c/glean/
  • Model: enterprise search + AI agents over existing content (permission-aware crawl/index). It is fundamentally a read layer over what already exists — no native mechanism found in research for capturing a fresh human answer as a new governed artifact, nor for routing an unanswered question to a specific human and writing the resolution back. Glean’s own marketing frames it as agent-building platform, not a Q&A-to-artifact loop closer.
  • Verdict: best-in-class retrieval/search substrate, does not itself close the human-answer -> durable-artifact loop; would need to be paired with something else (or custom agent) for that half.

3. Guru

  • 2026 update: Federated Search indexes Google Drive/Box/Confluence without migration (2026). Slack MCP integration launched March 2026 for querying live conversations. Source: https://techplustrends.com/guru-vs-glean-2026-enterprise-ai-knowledge-tools/ (2026); Guru product pages
  • Pricing shifted to an AI-credit/resolution model in 2026 rather than pure per-seat.
  • This is the closest built-in “full loop” product found. Concrete mechanics, verified against Guru help docs:
    • Any Slack message can be turned into a Guru Card directly from the message menu (Creator role) — this is the “human answer -> durable artifact” step, manual but one click, no separate authoring tool needed. Source: https://help.getguru.com/docs/creating-and-adding-to-guru-cards-in-slack
    • Verification workflow: card creator picks a Verifier + a verification interval (default: creator, every 3 months). Card turns yellow after 90 days idle, auto-notifies verifier via Slack/email, one-click re-verify — this is the missing “durable” / anti-decay half that Stack Internal lacks. Source: https://www.getguru.com/features/verification ; https://help.getguru.com/docs/verifying-guru-cards-in-slack
    • Reported outcomes: McKinsey-cited 35% search-time reduction (~7 hrs/week/ employee); Forrester TEI 20-25% productivity gains (vendor-commissioned studies — treat as directional, not independent). Source: techplustrends.com / eesel.ai Guru reviews 2026
  • Gap: verification is a manual SME chore on a timer, not an automatic “was this ever asked again and answered differently” reconciliation. No evidence found of an automatic “unanswered question routed to a human, human answers, system drafts and files the card” fully autonomous chain — a human still has to proactively convert the Slack message.
  • Verdict: closest existing product to the full loop, but the write-back step is human-initiated (one click) rather than agent-initiated; the decay check is scheduled, not triggered by contradiction detection. Good reference model, not a drop-in solve for an agent-population where 64 agents would need to trigger card creation themselves.

4. Notion AI (Q&A / Enterprise Search)

  • “Ask AI” / Q&A cites source pages when answering; grounded, not synthetic- only. 2026: Enterprise Search extends Q&A across connected apps (Slack, Jira, GitHub, Google Drive). Notion Agents can now do up to ~20 min of autonomous multi-step work across pages. Source: https://www.notion.com/help/guides/get-answers-about-content-faster-with-q-and-a ; https://fazm.ai/t/notion-ai-features-2025-2026 (2026)
  • Pricing: full AI stack bundled into Plus ($10/user/mo) and Business ($15/user/mo) as of 2026, no separate AI add-on charge.
  • No evidence of automatic escalate-to-human-then-write-back-as-new-doc loop; it’s read/synthesize over docs a human already wrote, plus agents that can create documents on request — but nothing found that specifically detects “this question has no good source,” routes to a named human, and files the human’s reply as a new canonical doc automatically.
  • Verdict: strong grounded-answer UX, no purpose-built unanswered-question routing/write-back loop; would need custom automation on top.

5. Atlassian Confluence / Rovo (AI answers)

  • Rovo bundled free into all paid Confluence Cloud plans: Premium/Enterprise from April 2025, Standard from October 2025. Rovo Chat/Agents cost 10 credits/request, Deep Research 100 credits/request; Rovo Search itself is free/uncredited. Source: https://www.atlassian.com/software/confluence/ai ; https://www.eesel.ai/blog/atlassian-brings-an-ai-assistant-to-jira-and-confluence (2026)
  • Team ’26 (May 2026) shipped the largest Rovo update set; “Remix” (open beta early 2026) converts page sections into charts/infographics/timelines.
  • Rovo Chat is a read-time assistant across Confluence/Jira/Slack/apps — no evidence of a native “question unanswered -> ping human -> human answer becomes new Confluence page automatically” flow. Confluence has always supported humans manually turning a Slack thread/Jira comment into a page, but that’s process, not product feature.
  • Verdict: enterprise-grade retrieval/chat layer over existing wiki content; loop-closing write-back is not a shipped capability as of 2026-07.

6. Docs-answering bots: kapa.ai, Inkeep, RunLLM — DO route to human & (partially) write back

This is the strongest cluster of evidence that vendors are explicitly building toward the full loop, though maturity/company size varies a lot.

kapa.ai

Inkeep

  • Escalation: hands off to a human “with full conversation history, knowledge sources consulted, and customer’s inferred intent” when not confident or when the ask requires account-specific action.
  • Closest to genuine automated write-back found in this research: Inkeep “can auto-reply to new support tickets, emails, GitHub issues, or forum posts only when confident… detects gaps from real signals, drafts updates with sources, and routes them through your PR review” — i.e. documentation-gap detection -> AI drafts the fix -> human approves via PR. This is a genuine (if human-gated) instance of “durable artifact” creation triggered by the Q&A signal itself, not just Slack-message-to-card copy/paste. Source: https://inkeep.com/use-cases/documentation-leaders (2026)
  • No independent funding/customer-count figures were pulled in this pass; treat scale claims from inkeep.com as vendor-sourced pending 3rd-party confirmation.

RunLLM

  • Support handoff: Discord/Slack threads get copied to an internal team channel; answers from there propagate back into the original thread; users can explicitly “pass unanswered questions to a real human.”
  • Feedback loop: admin downvote in Slack triggers a feedback form to immediately “tutor” the bot; company states it “gets smarter with every investigation” via reinforcement-learning-style feedback capture. Source: https://docs.runllm.com/ ; https://docs.runllm.com/release-notes/
  • Company status: bootstrapped, no confirmed outside funding as of the most recent (2025-09-02) source; ~$1.8M ARR, ~$5.3M valuation estimate, ~16 employees. Small, unproven at large scale. Source: https://getlatka.com/companies/aqueducthq (updated 2025-09-02)
  • Verdict on this cluster: kapa/Inkeep/RunLLM are the vendors explicitly marketing the exact loop asked about (unanswerable -> human -> artifact -> next-asker self-serves). Inkeep’s PR-gated auto-draft is the most literal match found anywhere in this research. All three are small, VC- or bootstrap-scale companies (not Atlassian/Notion/Glean scale) — real capability, unproven durability/vendor-risk profile.

7. Two more loop-closing product entrants (found opportunistically)

Question Base (questionbase.com)

Slite (“self-maintaining knowledge base”)

  • Slite Agent ships 2026-06-10 on the Pro plan ($20/user/mo annual): an agent that “detects when documentation has drifted from reality, proposes the fix, and routes every change through human approval before it becomes truth” — contradiction/staleness detection is the differentiator vs Guru’s timer-based verification. Source: https://slite.com/changelog/the-self-maintaining-knowledge-base (2026); https://slite.com/blog/slite-announcing-self-maintaining-knowledge-base
  • Framed explicitly around agent-readiness: cites McKinsey 2025 enterprise survey — 23% of enterprises scaling agents in at least one function, 80% of those cite shaky data/context as the roadblock. Directly relevant framing for a 64-agent setup.
  • Verdict: Slite’s drift-detection-triggered write-back is closer to automatic (agent proposes on its own initiative, not just on a timer) than Guru’s; unclear yet (product 5 weeks old at research time) how it performs at scale or how well drift-detection actually works in practice. No 3rd- party review evidence yet — too new.

8. Backstage TechDocs — real-world adoption pain (VERIFIED via GitHub RFC)

  • Not “does TechDocs work” but “what does it cost to run it”: requires external infra (S3/GCS + IAM), a dedicated CI/CD pipeline to render Markdown->static HTML, and for production-grade search a self-managed Elasticsearch/OpenSearch cluster. “Organizations satisfied with their self-hosted setup typically dedicate 3 to 12 full-time engineers to the platform.” Source: https://roadie.io/blog/backstage-how-much-does-it-really-cost/ (vendor-competitor source, directionally consistent w/ GitHub RFC below)
  • Confirmed via primary source (GitHub RFC #33990, backstage/backstage, accessed 2026-07-18): TechDocs’ underlying engine is at real risk.
    • MkDocs itself: “effectively unmaintained since its last release (v1.6.1, August 2024).”
    • Material for MkDocs (the theme TechDocs depends on) “entered maintenance mode as of November 2025” — critical bug/security fixes only, through at least November 2026, no new features.
    • A proposed MkDocs 2.0 rewrite would drop the plugin system entirely, breaking “300+ plugins the ecosystem depends on” including TechDocs’ own mkdocs-techdocs-core integration plugin.
    • Backstage’s proposed fix is migrating to Zensical, a successor built by the Material for MkDocs team, chosen for native mkdocs.yml compatibility over Docusaurus/Sphinx. Source: https://github.com/backstage/backstage/issues/33990 (RFC, open as of 2026-07-18)
  • Separate: “Port.io” and Medium pieces argue “Backstage is dead,” platform engineers “spend their weeks fixing broken TechDocs pipelines instead of building their platform” — treat as competitor-authored (Port sells a Backstage alternative) but directionally corroborated by the primary-source RFC above. Source: https://www.port.io/blog/backstage-is-dead (vendor blog, bias noted)
  • Verdict: TechDocs adoption pain is real and dated precisely — its whole foundation (MkDocs/Material) is heading into unmaintained territory on a named clock (Nov 2026), independent of any Q&A-loop question. Would be a risky foundation to build new internal docs tooling on right now without planning for the Zensical migration.

9. ADR tooling health

  • adr-tools (npryce): 5.6k stars, 631 forks — but latest release is v3.0.0 from 2018-07-25. No archive banner found, but no evidence of activity in the 2025-2026 window either. Treat as de facto unmaintained / stable-but-frozen (bash scripts implementing the Nygard ADR format; the format itself needs no updates, which is likely why nobody’s forced to touch the tool). Source: https://github.com/npryce/adr-tools (accessed 2026-07-18)
  • log4brains (thomvaill): latest release v1.1.0, 2024-12-17; 1.5k stars, 112 forks, 26 total releases, 149 commits on develop, 45 open issues, CI configured, no archive/deprecation notice. This is a genuinely more actively maintained project than adr-tools (real releases into late 2024, issue activity through Jan 2026 per issue-tracker dates seen in search), though the ~7-month gap since last release (Dec 2024 -> Jul 2026) plus 45 open issues warrants a “watch, don’t bet the farm” read. Source: https://github.com/thomvaill/log4brains (accessed 2026-07-18)
  • Verdict: ADR tooling is a mature-and-basically-done space (the format is simple markdown+numbering; tool churn isn’t needed) rather than an actively innovating one. Neither tool has any Q&A-loop or AI capability — ADRs remain pure “the two humans write a markdown file” practice, no self-service Q&A layer exists in this space at all.

10. DDD ubiquitous-language / glossary tooling — Contextive

  • Contextive (dev-cycles/contextive): IDE extension (VS Code, IntelliJ, Visual Studio, Neovim, Helix) that surfaces glossary term definitions via hover/autocomplete wherever a term appears — code, comments, config, markdown docs. Glossary stored as a YAML file in-repo, supports aliases.
  • General glossary practice (not tool-specific): consistent guidance across ddd-practitioners.com, martinfowler.com/bliki, SSW.Rules, RST Software — “the output of the Ubiquitous Language practice should be a glossary of terms & concepts,” commonly landing in a wiki page-per- subdomain (Confluence pattern cited) with links out to deeper docs; the language “isn’t something created once — it evolves.” No tooling beyond Contextive was found that operationalizes this as a live, code-adjacent, IDE-surfaced artifact; everything else is “write it in the wiki and hope people read it,” i.e. the same durability problem every other wiki has. Source: https://ddd-practitioners.com/home/glossary/ubiquitous-language/ ; https://martinfowler.com/bliki/UbiquitousLanguage.html
  • Verdict: Contextive is small but real and is the only glossary-specific tool found that ties directly into where terms are actually used (code/ docs) rather than living in an isolated wiki page nobody visits — but it has zero Q&A/answer-capture mechanics; it’s read-only surfacing of human-curated definitions, not a loop-closer.

Cross-cutting answer to the research question

No single existing product closes the full loop (question -> human answer -> durable artifact -> next asker self-serves) end-to-end and autonomously at the scale this project cares about (many agents, few humans). The closest matches, ranked:

  1. Inkeep — genuine gap-detection -> AI-drafts-the-fix -> human-approves- via-PR chain; smallest company of the “big three” (kapa/Inkeep/RunLLM), scale/durability unproven independently.
  2. Guru — the most field-proven “convert human answer to durable card + scheduled re-verification” loop, but the write-back step is manual (one click) and decay-checking is timer-based, not contradiction- triggered; largest install base/track record of this list.
  3. Slite Agent (2026-06-10, five weeks old at research time) — contradiction/drift-triggered write-back is architecturally closer to “autonomous” than Guru’s timer, but essentially unproven in the field.
  4. kapa.ai / RunLLM — solid “route unanswered to human” half of the loop, weaker/less-verified “durable artifact” half (kapa: signals gaps for humans to fix manually; RunLLM: propagates answer back into the same thread, not clearly into a persistent searchable KB entry).
  5. Question Base — pitches the exact loop, interesting pattern reference, but an 18-month-old pre-seed startup — vendor risk too high to lean on for an internal-only project without a fallback plan.

Everything else surveyed (Glean, Notion, Confluence/Rovo, Stack Internal/Teams) is a read/search layer over content humans already wrote, with Confluence/Notion/Rovo adding “AI chat over existing docs” and Glean adding “AI agents over everything,” but none of the big, proven platforms were found to have a native, verified “unanswerable question routes to a named human, human’s answer becomes a new governed artifact, staleness is actively detected and reconciled” pipeline. That specific capability lives almost exclusively in the smaller docs-support-bot vendor category (kapa/ Inkeep/RunLLM/Question Base) and in Guru/Slite’s card-verification / drift-detection mechanics — none of which is bulletproof, all of which are either small-vendor risk or (Slite) too new to have a track record.

For a 2-human + 64-agent internal setup, the implication is: buy or copy the pattern (Slack/chat message -> one-click promote to durable card with a named verifier and a decay/drift check), not any single vendor wholesale — and expect to build the “agent notices it can’t answer, tags a human, captures the human’s reply as a new artifact automatically” step yourself, since even the vendors explicitly selling that story (Inkeep excepted, PR- gated) mostly still require a human to manually promote the answer.