Knowledge base

Configuration

m8b:
  embeddings:
    baseUrl: http://vllm.example.internal:8000/v1   # OpenAI-compatible endpoint, ending in /v1
    model: embedding-model                          # exact embedding model ID
  knowledgeBase:
    bootstrap:
      enabled: true
      mode: if-missing       # keep an existing non-empty index; "always" rebuilds it
      timeoutSeconds: 3600   # Job deadline, separate from the Helm timeout

• The embedding endpoint needs network egress: m8b.ai.egress covers it when it is the same server as the chat endpoint, otherwise set m8b.embeddings.egress (or podEgress for an in-cluster server).
• If the embedding endpoint needs a different key than the chat endpoint, add it to the m8b-runtime Secret as AI_EMBEDDING_API_KEY.
• The index records the embedding configuration. After changing the embedding model or endpoint, rebuild it.
• Embedding prefixes and input types, described in the bot's configuration reference^[1], are set with m8b.embeddings.queryPrefix, documentPrefix, queryInputType and documentInputType. The chart rejects the corresponding variables in m8b.extraConfig, because it manages them.

Build, rebuild

Follow GETTING-STARTED.md^[2], step 11: stop the bot, run stage=index, save the Job logs, delete the Job, then stage=run. The bot and the Job must never use the volume at the same time, which is why the bot is stopped explicitly first: do not upgrade a running bot straight into index and expect Helm to wait for its Pod.

To force a rebuild of an existing index, run the same step 11 with --set m8b.knowledgeBase.bootstrap.mode=always on the stage=index command only. The previous index is backed up next to the new one, and restored if the rebuild fails. The final stage=run command, without that flag, returns to if-missing.

Good to know

• A bot restart does not reindex: a new bot Pod loads the existing index from the volume.
• The index file may only be written at the end; no visible progress on disk is not a sign of a hang. Follow the Job log instead.
• On failure, the Job and its log are kept and the bot stays stopped. Read the log, fix the cause, delete the Job (Helm does not remove a Job created by a failed revision), then run stage=index again.
• Do not helm rollback to an index revision: that would stop the bot and start an indexer again.
• A Ready bot does not prove the index works: finish with a Slack question that only the documentation answers.
• The index Job has no Kubernetes API credentials.