Give retrieved fragments an identity before changing the model
Contextual Retrieval can restore clues lost during chunking. Keep source identity, citations and access rules intact before judging whether enrichment helps.
By 
Paul. AI and agentic engineering
A passage can contain the right number and still be unusable evidence. Without the company, reporting period or document revision, a retrieval system may select it for the wrong question. Before replacing the model that writes an answer, establish whether the selected fragments still identify what they describe. Restoring missing context is a concrete retrieval intervention; its value depends on preserving the original evidence.
Looking back from November 2026, Anthropic's 19 September 2024 Contextual Retrieval article supplies a useful mechanism. It prepends a short, document-informed explanation to each chunk before embedding and BM25 indexing. That changes what search can recognize. It does not establish that the explanation is true or that the requester may read the source.
Separate the passage from its description
Consider a proposed assistant searching successive commercial reports. A paragraph says that revenue increased, while the heading identifying the business and quarter sits outside the chunk. Searching the paragraph alone can miss the relevant business or find a similar sentence from another reporting period. This is an illustrative design problem, not a result observed in an AL+ system.
A contextual description could restore those missing clues. Store it separately from the original passage, even if they are combined for indexing. Keep a stable document identifier, revision and source location with both. A reviewer should be able to see which words came from the document and which were generated during preparation.
This separation determines what an answer can cite. A link to an enriched chunk should open the underlying passage in its original context. If the explanation says a policy applies to all customers but the source limits it to one contract, the explanation must not become its own supporting evidence. The original qualification controls the answer.
Where reliable metadata already supplies a title, date or customer identifier, use that rather than asking a model to infer it. Generated context is most interesting when meaningful relationships are present in the document but absent from the fragment. Repeating known fields through generation introduces another opportunity for error without necessarily adding information.
Read the percentage as a retrieval result
Anthropic reported a reduction in top-20 retrieval failure from 5.7% to 2.9% with contextual embeddings plus contextual BM25, a relative reduction of 49%. With reranking it reported 1.9%, or 67% relative reduction. The displayed aggregate used Gemini Text 004 embeddings across its evaluated domains; the reranking experiment used Cohere. The metric was one minus recall at 20. These are vendor findings under stated conditions, not a 49% improvement in answer accuracy.
That distinction changes the purchasing decision. Finding the right evidence is necessary for many answers, but a generator can still misread it. Conversely, retrieving twenty chunks can count as success while providing a confusing mixture of current and obsolete passages. A retrieval score and an accepted business answer deserve separate records.
Neither percentage forecasts results for Romanian documents, scanned contracts or your access-filtered corpus. Different extraction quality, abbreviations and questions can change which failure dominates. The source supports trying the intervention where chunk identity is lost. It does not justify promising the same gain in a proposal.
Enrichment inherits the source's restrictions
The preparation stage needs authorization too. If a service reads a restricted whole document to explain a public-looking paragraph, its generated description can carry confidential information into that paragraph's index entry. Making the raw fragment accessible does not make the whole-document explanation safe to share.
A defensible design keeps enrichment within the source document's access boundary. Restrict candidate selection before sending material to the answer model. Preserve those restrictions in search indexes, reranking inputs and caches. A citation should also respect access at the moment it is opened; an identifier is not a public download capability.
Revision handling follows the same reasoning. Replacing a source must replace or invalidate descriptions derived from its former contents. Keep enough provenance to rebuild the index and trace an answer to the version actually supplied. A more searchable obsolete policy can be more harmful than a difficult-to-find one because it appears convincingly relevant.
Isolate the missing-context hypothesis
A proposed comparison should hold the questions, source revisions, access identity, chunk boundaries and answer model fixed. Compare raw chunks with metadata-only enrichment and with document-generated descriptions. Keep the retrieval budget comparable. Otherwise an apparent gain could come from returning more material or changing several components together.
Inspect misses before counting successes. Did the relevant source enter the candidate set? Did ranking discard it? Did the final answer contradict a qualification that was already present? Only the first two are principally retrieval problems. If the required statement never existed, enrichment should not manufacture an answer.
Include questions that distinguish almost identical revisions and questions the permitted documents cannot answer. Check citation support separately from fluent prose. These are suggested evaluation cases, not experiments executed for this article. The business consequence being tested is fewer unsupported answers and less correction work, with preparation and runtime costs counted separately.
A whole document can remain the better unit
The strongest objection is that a small, stable collection may not need fragment search at all. Supplying an authorized whole document can preserve relationships without an enrichment pipeline. Larger chunks or reliable headings may also solve the observed problem more cheaply. Contextual Retrieval earns its preparation and revision work only when those simpler options are insufficient.
For an AI assistant project, request one traced failure before approving a model replacement: the question, selected source revision, retrieved passage and unsupported conclusion. If the passage lost its identity, repair that boundary and compare again. Accept enrichment when it helps retrieve usable evidence while keeping the source, its limits and its permissions visible.