RAG Failure Modes

Seven failure classifications documented in production deployments. RAG Axis makes each one a named, catchable type.

Missing Content

The correct document does not exist in the corpus. The model fabricates a plausible answer. RAG Axis response: EmptyRetrievalError with corpus coverage signal.

Missed Top-K

The correct chunk is indexed but scored too low to survive the top-k cutoff. The system behaves as if the data is missing. RAG Axis response: BelowThresholdRetrieval with highest_score exposed.

Out-of-Context Chunk

The chunk is retrieved but the chunking algorithm severed semantic dependencies. The LLM misinterprets the isolated fragment. RAG Axis response: Chunk provenance (position, parent doc) mandatory via I1.

Not Extracted

The relevant context is retrieved and injected but the LLM ignores it. Lost-in-the-middle syndrome. RAG Axis response: Context ordering strategy in ragaxis.context.

Wrong Format

The LLM ignores output schema constraints. Downstream services crash on unparseable payloads. RAG Axis response: Schema-enforced output parsing in ragaxis.generation.

Incorrect Specificity

The response is factually accurate but misaligned with user intent. RAG Axis response: Intent classification hook in query processing.

Incomplete Generation

The response addresses only part of a multi-intent query. RAG Axis response: Multi-intent decomposition in query processing.

Silent Failures

The most dangerous class. HTTP 200, normal latency, confident but wrong answer. Caused by relevance drift, knowledge staleness, and embedding model mismatch. RAG Axis response: corpus_version, content_age_days, EmbeddingModelMismatchError, and ScoreCollapseWarning are all mandatory signals, never optional.