vector-search-ops

# vector-search-ops — Vector Search Operations Covers day-to-day operations for the `resources/vector_search.yml` component scaffolded by AgentOps Stacks: endpoint health, index sync, retriever testing, and DLT pipeline changes. ## When to use - "my retriever isn't returning results" - "how do I re-sync the index after adding documents?" - "check vector search status" - "the DLT pipeline failed" - "test the retriever" - "how do I update the embedding source?" ## Key resources in the project | File | Purpose | |---|---| | `resources/vector_search.yml` | DAB resource: VS endpoint + delta-sync index | | `src/components/retriever/data_pipeline.py` | DLT pipeline: raw docs → chunked docs → VS index | | `src/agents/<name>/graph.py` | Retriever wiring (`get_relevant_chunks`, `retriever_node`) | The index is named `<catalog>.<schema>.<project>_vs_index` and syncs from `<catalog>.<schema>.<project>_chunked_docs`. --- ## Operations ### 1. Check endpoint and index status ```python from databricks.sdk import WorkspaceClient w = WorkspaceClient() # Endpoint ep = w.vector_search_endpoints.get_endpoint("<project>_vs_endpoint") print(ep.endpoint_status) # ONLINE / OFFLINE / PROVISIONING # Index idx = w.vector_search_indexes.get_index( index_name="<catalog>.<schema>.<project>_vs_index" ) print(idx.status) # ONLINE / PROVISIONING / FAILED print(idx.index_url) # endpoint URL ``` Or via CLI: ```bash databricks vector-search endpoints get --name <project>_vs_endpoint databricks vector-search indexes get-index --index-name <catalog>.<schema>.<project>_vs_index ``` ### 2. Trigger a manual sync The index uses `pipeline_type: TRIGGERED` — it does NOT auto-sync on table writes. Trigger it explicitly after new documents are ingested: ```python from databricks.sdk import WorkspaceClient w = WorkspaceClient() w.vector_search_indexes.sync_index( index_name="<catalog>.<schema>.<project>_vs_index" ) print("Sync triggered — check status with get_index()") ``` Wait for sync to complete before testing: ```python import time while True: idx = w.vector_search_indexes.get_index(index_name="<catalog>.<schema>.<project>_vs_index") if idx.status.detailed_state in ("ONLINE", "ONLINE_NO_PENDING_UPDATE"): break print(f" {idx.status.detailed_state} — waiting...") time.sleep(15) ``` ### 3. Test the retriever Run a quick retrieval test without starting the full agent: ```python from databricks.sdk import WorkspaceClient from databricks_langchain import DatabricksVectorSearch import os w = WorkspaceClient() vs_index = "<catalog>.<schema>.<project>_vs_index" retriever = DatabricksVectorSearch( index_name=vs_index, columns=["content", "doc_uri"], workspace_client=w, ).as_retriever(search_kwargs={"k": 5}) docs = retriever.invoke("your test query here") for doc in docs: print(f"[{doc.metadata.get('doc_uri', '?')}]\n{doc.page_content[:200]}\n") ``` ### 4. Ingest new documents New documents go into the UC Volume at `/Volumes/<catalog>/<schema>/artifacts/raw_docs/`. After uploading: 1. Run the DLT pipeline (`data_pipeline.py`) to update the `_chunked_docs` table 2. Trigger the VS index sync (step 2 above) ```bash # Upload a file to the volume databricks fs cp my_doc.pdf dbfs:/Volumes/<catalog>/<schema>/artifacts/raw_docs/ # Run DLT pipeline via bundle databricks bundle run -t dev data_pipeline # Then trigger sync (Python snippet above or SDK) ``` ### 5. Update the DLT pipeline (chunking logic) Edit `src/components/retriever/data_pipeline.py`. The two DLT tables are: - `<project>_raw_docs` — reads from the UC Volume - `<project>_chunked_docs` — apply chunking here Common changes: - **Different source format** — swap `cloudFiles.format` in `raw_docs()` - **Custom chunking** — replace the body of `chunked_docs()` (LangChain splitters, unstructured.io, etc.) - **Add metadata columns** — extend `.select()` and `columns_to_sync` in `vector_search.yml` After editing, redeploy: ```bash databricks bundle deploy -t dev databricks bundle run -t dev data_pipeline # then trigger sync ``` ### 6. Change the embedding model In `resources/vector_search.yml`, update the `embedding_model_endpoint` variable or the `embedding_model_endpoint_name` field on the index. The built-in default is `databricks-gte-large-en`. **Note:** Changing the embedding model requires rebuilding the index from scratch. Delete and recreate the index (or undeploy + redeploy the bundle resource). ```bash databricks bundle destroy -t dev # removes VS endpoint + index databricks bundle deploy -t dev # recreates with new model databricks bundle run -t dev data_pipeline # then sync ``` ### 7. Increase `k` (number of retrieved chunks) In `src/agents/<name>/graph.py`, `get_relevant_chunks(query, k=5)` — increase `k` for broader retrieval, decrease for more precise context. Also adjustable per-call: ```python retriever.as_retriever(search_kwargs={"k": 10}) ``` --- ## Troubleshooting | Symptom | Likely cause | Fix | |---|---|---| | Index stuck `PROVISIONING` | VS endpoint not ready | Wait; endpoint creation takes 5–10 min on first deploy | | Empty results | Index not synced after ingestion | Trigger sync (step 2) | | Empty results | Chunked docs table empty | Run DLT pipeline first | | `FAILED` index status | Embedding model endpoint down | Check `databricks-gte-large-en` is available in your workspace | | Retriever returns irrelevant chunks | `k` too high or chunking too coarse | Reduce `k`, improve chunking in `data_pipeline.py` | | `403` on `sync_index` | App SP lacks permission | Grant `SELECT` on the chunked_docs table to the app SP | | Column not returned in results | Column not in `columns_to_sync` | Add it in `resources/vector_search.yml` and re-sync | ## Environment variables (set in `app.yaml`) ```yaml env: - name: CATALOG value: <catalog> - name: SCHEMA value: <schema> ``` The retriever in `graph.py` reads `os.environ["CATALOG"]` and `os.environ["SCHEMA"]` to build the index name at runtime.