Skip to content

ADR-0002: Native Bot KB Strategy — COS + WeCom Knowledge Sets, Lexiang Off the Critical Path

  • Status: accepted
  • Date: 2026-09-26
  • Deciders: Huashan Wang
  • Supersedes: ADR-0001 (knowledge-source portion only — see below)

Context

ADR-0001 deployed the OVES Product Expert bot on WeCom Smart Robot with its knowledge set linked to the Lexiang Articles space (浩勉 tenant) via MCP. The Lexiang linkage caused recurring friction:

  • Custom MCP integration glue was required to push knowledge entries.
  • The doc → Lexiang publish pipeline failed repeatedly (plantuml_markdown dependency missing since 2026-08-11).
  • The tenant adds a vendor, a billing surface, and a version dependency that OVES does not control.

Meanwhile the organization already operates COS + git repositories as the single knowledge SPOT (tencent-cloud ADR-0015). WeCom Smart Robot's native 知识集 (knowledge set) accepts direct file upload (up to 1,000 files per set) and provides built-in RAG with an explicit fidelity control:

Reply mode Behavior
仅根据知识集回答 Strict — answers only from the KB, refuses otherwise
联网搜索获取最新知识回答 Bot searches the web for fresh knowledge
Default (neither) KB + model native knowledge, most creative

Lexiang's 2026 upgrades (Agentic KB, zero-code WeCom/Feishu/DingTalk connectors, ADP one-click WeCom publishing) were reviewed on 2026-09-26 and assessed as not required for current scope.

Decision

  1. Bot knowledge sets are built from staged file packages, not from Lexiang: git repositories + COS → curated Markdown → 知识集 upload. One product scope per bot, staying well under the 1,000-file cap.
  2. Lexiang is off the critical path for all WeCom bots. No new content is pushed to Lexiang for bot consumption.
  3. Fidelity starts strict: every new bot launches in KB-only mode (仅根据知识集回答) and may be loosened per bot after pilot review.
  4. A sync script (repos/COS → staged Markdown → knowledge-set files) is owned in this repo and run on a cadence (weekly by default).

Lexiang re-entry triggers

Revisit Lexiang only when at least one of these becomes real:

  1. A bot must execute tasks (write to systems), and Lexiang/ADP's Agentic tooling beats the alternatives evaluated at that time.
  2. Real-time WeCom → KB sync (capturing chat/doc knowledge automatically) is required.
  3. Multiple bots need one governed, deduplicated, versioned shared KB with review workflows.
  4. A single bot's knowledge scope outgrows the 1,000-file-per-set model.

Consequences

Benefits

  • Zero new vendors, tenants, or billing surfaces.
  • Knowledge source of truth stays COS + git (consistent with tencent-cloud ADR-0015).
  • The fidelity dial is native, per-bot, and reversible without redeployment.
  • The broken doc → Lexiang publish script is no longer a bot blocker.

Trade-offs

  • Knowledge updates are batch (cadence-driven), not real-time.
  • 知识集 upload is admin-console work (or future API automation); no incremental sync engine is provided.

Relation to ADR-0001

ADR-0001's bot-platform decision (WeCom Smart Robot as the Q&A channel) is unchanged. Only its knowledge-source linkage (Lexiang Articles space via MCP) is superseded by this ADR. The Lexiang MCP connection details in ADR-0001 are retained for historical reference and the parked doc-publish fix.