Vector search is good at semantic similarity, but business context also depends on exact identity, authority, date, relationship, and permission. A close passage can be operationally wrong because it belongs to another company, an obsolete policy, a draft, or a restricted source. Reliable retrieval therefore combines dense search with lexical, structured, graph, and policy-aware routing.
Know which questions similarity cannot answer
Embeddings represent patterns in language. They can connect a question about cancellation with a policy that uses different wording. They do not inherently know that one customer identifier differs by a character, that a decision superseded an older note, or that a subsidiary has separate rules.
Similarity also does not establish authority. A frequently repeated blog post may rank above a signed policy. A draft proposal may resemble the query more than the live product catalogue. A vector score says how close representations are, not which record the business has authorized for the claim.
Permissions cannot be repaired after unrestricted retrieval. If a restricted fragment enters the model context, filtering the final sentence is too late. Candidate sources must be constrained before content is read. The public article on company memory beyond vector search discusses this broader architecture without proving retrieval quality for another system.
Freshness is another external property. Re-embedding can update an index, but it does not decide which effective date applies or whether a delayed source sync makes the result unsafe. That logic comes from metadata, source contracts, and domain rules.
Route with exact, structured, and policy signals
Begin with authenticated identity and purpose. Use these to choose eligible entities, sensitivities, and source families. Then apply exact lookup for customer, contract, product, or ticket identifiers. Lexical search handles names, codes, and rare terms that dense retrieval may blur.
Use structured filters for effective dates, status, region, language, owner, and document type. Graph relations can connect a subsidiary to the policy that actually governs it or a product to its approved catalogue. Dense retrieval then ranks semantically relevant material inside the permitted candidate set.
After selection, hydrate the authoritative source and verify its version. If evidence conflicts or the source is unavailable, route to the defined owner instead of broadening silently. This sequence makes semantic search a useful component rather than an accidental policy engine.
The trust checklist for confident AI failures can help teams inspect plausible wrong answers. It does not certify a hybrid retrieval pipeline or its permission controls.
Evaluate business mistakes, not generic relevance
Create test cases from costly retrieval errors: parent policy returned for a subsidiary, old price selected over current catalogue, draft treated as approved, exact client confused with a similar name, restricted source exposed, and missing evidence replaced by a plausible passage.
Label the acceptable source set, blocked sources, required version, and expected refusal behavior. Review retrieval before generation as well as the final answer. Otherwise a model can hide a poor candidate set by producing smooth language from incomplete evidence.
Measure correct source selected for the business question, unauthorized candidates removed, stale records rejected, conflicts surfaced, and recovery after source change. Report severe boundary failures separately. One cross-entity exposure should not disappear inside an average relevance score.
Keep evaluation queries and source versions reproducible so routing changes can be compared fairly.
The verified completion memory release is public evidence of maintained tooling and test work, not a universal benchmark for this retrieval design.
Frequently Asked Questions
A retrieval architect designs the routing with domain owners, who supply the authority, date, entity, and permission rules for their records.
Use exact identifiers, metadata filters, keyword matches, graph relations, source versions, and permissions. The combined trace should reveal which signal admitted, ranked, or blocked each candidate.
It can blur exact identifiers, rank a draft above an approved record, miss effective dates, cross entity boundaries, or retrieve material the user is not allowed to read.
Measure correct-source selection, exact-identifier handling, rejection of stale and unauthorized records, conflict detection, safe refusal, and recovery after a source changes.
AI4SALE has built bounded retrieval and source-backed memory mechanisms and publishes their contracts, without claiming unmeasured customer outcomes. A hybrid routing and evaluation plan can be developed through the exact planned AI4SALE enterprise AI search service after its publication gate.