Knowledge and Search

Knowledge and search

Open PlayWeld > Plan & Collaborate > Knowledge & Canon; its tab is Knowledge. Retrieve Project material with citations, then inspect original files before making a decision.

Start lexically

1. Select the Project and Index status.
2. After saved edits, choose Reconcile. Wait for pending work/timestamps to settle.
3. Open Search > Search mode and filters, choose Lexical and source such as docs.
4. Search a distinctive phrase and inspect citation, path, excerpt and revision.

Lexical search requires no embeddings. Chunks can exist while canon count is zero: ordinary documents are not automatically canon records.

Add semantic retrieval

Semantic search needs an embedding profile/model and vector storage. Hybrid combines retrieval and can report degraded behavior. Test vector store checks storage, not the embedding provider or retrieval quality.

Choices include lexical-only, embedded LanceDB, SQLite exact-vector search and managed/local/remote Qdrant. SQLite exact search is not ANN. Lexical-only is the default; selecting storage does not install model weights.

Remote storage needs its explicit configuration/permission, HTTPS in remote mode and credential references. Embeddings/index metadata can leave the machine. Backend switching schedules reconciliation but retains old backend data; deletion there is separate.

Diagnose stale results

Save the source, inspect errors/pending work, reconcile, then query an exact new phrase lexically. Rebuild queues a full rebuild for a diagnosed issue. Do not delete canon to repair derived data.

See Knowledge configuration and recovery.