Databricks Context Engineer Associate flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Delta time travel

A Delta Lake feature that lets you query earlier table versions by version number or timestamp, giving reproducible snapshots of source data and chunk tables for evaluation and debugging.

Unity Catalog namespace

The three-level naming scheme catalog.schema.object used for tables, volumes, models, and functions, so context assets inherit centralized permissions, lineage, and discovery.

Foundation Model APIs

Databricks-hosted model endpoints billed per token, offering pay-per-token or provisioned throughput modes for chat and embedding workloads without managing serving infrastructure.

Provisioned throughput

A Model Serving mode that reserves dedicated capacity at a fixed tokens-per-second rate, giving predictable latency and cost for production context applications.

MLflow trace

A recorded execution of a GenAI request made of spans for retrieval, tool, and LLM calls, used to debug latency and inspect the exact context sent to the model.

Auto Loader

A Databricks ingestion source using cloudFiles that incrementally discovers new files and tracks them in a checkpoint, preventing duplicate rows on repeated runs.

Chunk overlap

The number of tokens shared between consecutive chunks, commonly 10 to 20 percent, preserving sentence context that would otherwise be split across boundaries.

Content hash deduplication

Computing a hash of normalized chunk text to detect identical or near-identical chunks, so duplicates are dropped before embedding and do not crowd retrieval results.

Quarantine table

A Delta table receiving rows that fail parsing or quality expectations, keeping bad records auditable instead of silently dropping them during ingestion.

Lakeflow expectations

Declarative data quality constraints on pipeline datasets, such as expect or expect_or_drop, that validate records at ingestion and report pass and fail metrics.

Delta Sync index

A Vector Search index that automatically syncs from a Delta table, so inserts, updates, and deletes in the source propagate to the index without manual writes.

Cosine similarity

A similarity metric measuring the angle between vectors, appropriate for normalized text embeddings because it ignores magnitude and compares direction only.

Hybrid retrieval

Combining dense vector similarity with keyword scoring such as BM25, then merging ranked lists, to capture both semantic matches and exact term matches.

Reciprocal rank fusion

A fusion method that sums one over a constant plus each document's rank across result lists, merging hybrid retrieval outputs without score normalization.

Metadata pre-filtering

Applying attribute filters such as date, document type, or access group inside the vector index query so top-k slots are filled with eligible chunks only.

Cross-encoder reranking

Scoring each query-document pair jointly with a cross-encoder model over the top candidates, improving precision before the final context is assembled.

Embedding model versioning

Storing the model name and version with every vector row, because vectors from different models are not comparable and mixed indexes return meaningless similarity.

Incremental re-embedding

Recomputing vectors only for new or changed chunks, detected via Change Data Feed or content hashes, instead of rebuilding the whole corpus each run.

Query versus passage prefix

An instruction prefix some embedding models require to distinguish search queries from stored passages; using the wrong prefix degrades retrieval quality.

Embedding batch size

The number of texts sent per embedding request, typically 100 to 1000; throughput depends more on endpoint concurrency than on a single large batch.

Asynchronous index build

Vector Search index creation and sync run in the background, so pipelines must poll index status before issuing queries or running evaluation.

Rolling conversation window

Keeping only recent turns verbatim plus a summary of older history, controlling token cost and preventing stale dialogue from diluting retrieved context.

AI Gateway

A Model Serving layer providing rate limits, payload logging, usage tracking, and guardrails centrally, so individual applications do not reimplement these controls.

Tool granularity

How narrowly agent tools are scoped; smaller typed tools produce clearer trace spans and more reliable orchestration than one broad catch-all tool.

Token budget allocation

Reserving fixed portions of the model context window for system prompt, retrieved context, history, and output, ensuring the request never exceeds the limit.

Low-confidence fallback

Returning a safe, explicit answer when retrieval yields no or weak matches, instead of letting the model answer from parametric memory alone.

Row filter propagation

Unity Catalog row filters and column masks on source tables carry into Delta Sync indexes, providing per-user retrieval scoping through existing permissions.

Embedding data sensitivity

Vectors derived from sensitive text can leak information, so they require the same access controls, retention, and deletion handling as the source documents.

Audit system table

The system.access.audit table recording endpoint and table access events, used to answer who queried which context asset for compliance reporting.

Recall at k

The fraction of labeled relevant chunks that appear in the top-k retrieval results, measuring whether the retriever finds the needed evidence at all.

Faithfulness metric

An LLM-judged score measuring whether claims in the generated answer are supported by the retrieved context, detecting hallucination in RAG outputs.

Golden evaluation set

A curated set of questions with reference answers and labeled supporting chunks, enabling separate scoring of retrieval quality and generation quality.