Vector Database¶
Vector Database - Similarity threshold distribution buckets thresholds into 6 bands (0.0-0.1, 0.1-0.3, ..., 0.9-1.0) so an over-strict threshold is visible without per-query inspection.
Purpose - RAG-side observability. Beyond simple "did the agent retrieve" questions, this tab surfaces retrieval parameters (top_k, similarity threshold) and per-DB returned-doc counts - the indicators that determine RAG quality.
When to look here¶
- "Did the agent actually use RAG?" - Traces with RAG count + Vector queries count.
- "Are queries slow against the configured store?" - Query latency p50 / p95 / p99.
- "Is
top_kbeing set sensibly?" - top_k distribution (a spike at 1 or above 50 usually indicates an agent miscalibration). - "Is the similarity threshold too strict?" - Similarity threshold distribution (everything in 0.9-1.0 → almost no recall).
- "Multi-store mix?" - DB systems donut (Chroma / Pinecone / SimpleVectorStore / etc.).
- "How many docs per query, on average?" - Avg returned docs by db.
Data source¶
Not trace spans: every vector-store operation is recorded in a memory-only operation log (VectorOperationLog) - up to 5000 operations plus the text of the last 20 queries, never written to disk and cleared on restart. Two kinds of search are left out: searches against the dynamic tool-discovery index (Agentic Chat's tool search) and the placeholder "list everything" query the Vector Database screen runs to fill its grid.
Controls¶
All dashboards share the Observability global settings - time window, refresh interval, custom range. Vector Database has no tab-specific controls beyond those.
KPI cards (four)¶
| Card | Shows | Source |
|---|---|---|
| Traces with RAG | Number of chat turns that included a vector query | TraceRecord.hasRag true count - set when a document-store query (RAG pipeline, document source, or chat attachment excerpt search) ran during the turn; tool-index searches do not count |
| Vector queries | Number of query operations (adds and deletes appear only in Operations / minute) |
Operation log, query count |
| Avg top_k | Mean of the requested top_k parameter |
db.vector.query.top_k attribute averaged |
| Avg query latency | Mean duration of query operations |
Operation log durations |
Charts (seven)¶
| Chart | Type | Reading |
|---|---|---|
| Operations / minute | Stacked bar (query / add / delete) | Add bursts during document ingestion; query bursts during chat. A pipeline with Multi-Query expansion issues one query per variant, so a single turn can show several |
| Query latency p50 / p95 / p99 | Multi-line, ms, query operations only |
Tail latency on remote vector stores can be ten-fold higher than SimpleVectorStore |
| top_k distribution | Histogram | Distribution of requested k values. Top-K is a per-pipeline setting, so a spread here means different RAG pipelines are in use rather than one value drifting |
| Similarity threshold distribution | Histogram, 6 bands ([0.0, 0.1) ... [0.9, 1.0]) |
Where the agent is anchoring relevance - strictness vs recall |
| DB systems | Donut by db.system |
One backend = single configuration; multiple = mixed stores |
| Avg returned docs by db | Horizontal bar | Combined with similarity threshold, this is the retrieval-recall signal |
| Recent searches | List, "last 20, kept in memory only" | One line per query: HH:mm:ss · N docs · N ms · <query text> (text cut at 200 characters). Empty state: No searches yet. The text never leaves memory and is gone after a restart |
Cross-references¶
- Vector Database (feature) - how documents are ingested and indexed
- Tutorial 3 - Index a Document - end-to-end ingestion walkthrough
- Tutorial 5 - Chat with RAG - consuming the indexed corpus
