Services / Retrieval
Retrieval systems that stay grounded after launch.
RAG is a source, permission, retrieval, evaluation and operations system. Generation is the final consumer of that evidence, not the architecture itself.
Source to answer
Keep the evidence path inspectable from ingestion to recovery.
- Control 01
Ingestion and source lifecycle
Name every source, sync mechanism and owner. Record what happens when a source is unavailable, replaced or removed so the index cannot silently outlive the system of record.
- Control 02
Parse, normalize and chunk
Preserve the structure needed for later decisions. Chunk size is tested against document type and query behavior; tables, headings, identifiers and versions are not flattened without a reason.
- Control 03
Metadata, retrieval and ranking
Combine lexical and semantic retrieval when the evidence supports it. Metadata filters, hybrid retrieval and reranking are evaluated separately so a ranking change has a visible regression signal.
- Control 04
Citations and provenance
Return source identity, location and version with the answer. Users need enough provenance to inspect a claim, while operators need enough trace to reproduce the retrieval path.
- Control 05
Tenant and source permissions
Apply identity and authorization before evidence reaches the model. Tenant boundaries and source-level permissions stay enforceable through indexing, caching, retrieval and generated output.
- Control 06
Freshness, deletion and reindexing
Define freshness targets, deletion propagation and reindex triggers. Stale documents are detectable, and a failed sync produces an operator-visible state rather than an apparently current answer.
- Control 07
Latency and cost
Budget ingestion, retrieval, reranking and generation independently. The service can then degrade deliberately, skip an expensive step or return search results when the full answer path misses its budget.
- Control 08
Observability, fallback and recovery
Track empty retrieval, low evidence coverage, permission rejection, stale-source use, model failure and end-to-end latency. Recovery includes replay, reindexing and a useful non-generative fallback.
Two failure domains
Measure retrieval before judging the answer.
Retrieval quality
Versioned queries test whether required evidence appears, whether forbidden evidence stays out and where relevant items rank. Empty or incomplete context is a retrieval failure even when the model writes a fluent response.
Generation quality
Answer evaluation checks support, citation accuracy, instruction adherence and safe refusal against the retrieved context. It does not give generation credit for evidence the retrieval layer never supplied.
Applicable evidence
Grounding and evaluation mechanics from documented work.
The linked career-survey work is evidence of structured extraction, grounding, evaluation and human review. It is not presented as a client RAG deployment.
- Evidence 01
Structured extraction, grounding, evaluation and human-review workflow for a high-stakes LLM system.
Read the engineering evidence
System choice
When RAG is the wrong system
The sources cannot support the decision
Retrieval cannot repair missing, contradictory or unowned source material. Source remediation or a narrower product claim comes before a generation layer.
The task requires exact lookup
Known identifiers, policy states and authoritative records are often safer through deterministic search or direct database access. Generated prose can remain outside the critical path.
Search already serves the user
If users need ranked documents and can interpret them directly, a strong permission-aware search interface may be more reliable, cheaper and easier to operate than answer generation.
Decision notes
Questions technical buyers ask
- Can you repair an existing RAG system?
- Yes, when the team can provide source access, representative queries and current failure evidence. The first step is to locate whether quality is lost in ingestion, retrieval or generation.
- Do you require a vector database?
- No. The query distribution and source structure determine whether lexical search, semantic retrieval, filters, reranking or a hybrid is justified.
- How do deletions reach generated answers?
- Deletion is traced through the source connector, index, caches and replay paths. The target propagation time and failed-deletion alert are part of the operating contract.