GOVERNED SEMANTIC SEARCH · BUILT WITH OVERLIFT

OverLift Semantic Search

Understand what people mean. Keep the answer defensible.

People rarely search with the exact vocabulary stored in a product, policy library, learning system, or research archive. OverLift Semantic Search recovers intent from imperfect wording, connects the user to useful evidence, and explains why a result deserves attention—without turning similarity into truth.

See how the search stays defensible Follow the product story
AzureGlossary semantic search and learning command center showing reviewed technical evidence and practice controls
Semantic search in a working product AzureGlossary demonstrates exact and semantic discovery over reviewed technical knowledge while official sources remain authoritative.
  1. 01
    Natural-language recoveryFind the right concept through scenarios, alternate terms, misspellings, and half-remembered language.
  2. 02
    Source-bound relevanceKeep results tied to reviewed evidence, permissions, currentness, and product-specific rules.
  3. 03
    Measured restraintEvaluate retrieval quality, duplicate results, citation correctness, abstention, and replay.
MARKETING

Search that feels intelligent—not like a filter with better branding.

Let people ask naturally, recover from imperfect wording, discover useful connections, and understand why a result deserves their attention.

People rarely use the exact vocabulary stored in a product catalog, knowledge base, policy library, learning system, or private research archive. OverLift Semantic Search meets them where they are: a question, a misspelling, a scenario, an outcome, or a half-remembered phrase.

Why it stands apart: the experience can be broad and forgiving at the proposal layer while remaining precise at the authority layer. The system can explain its match, preserve source identity, expose exclusions, and abstain instead of manufacturing certainty.

PRODUCT SIGNALSWhy users notice, trust, and return to the search experience
  1. 01Ask in ordinary language instead of memorizing product vocabulary
  2. 02Recover misspellings, alternate terms, and scenario intent
  3. 03Move from one result into connected concepts and useful next steps
  4. 04See the source, evidence, and reason behind an important match
  5. 05Receive a clear abstention when the corpus cannot support an answer
  6. 06Turn search into onboarding, learning, conversion, and product discovery
BUSINESS / EXECUTIVE

Turn scattered knowledge into a private, measurable product capability.

Reduce search time, support load, training friction, provider dependence, and data exposure while improving discovery and decision quality.

Semantic search becomes commercially useful when it improves a real journey: finding a product, resolving a support issue, learning a complex domain, locating an approved procedure, comparing evidence, or deciding what to do next.

Executive case: OverLift makes quality measurable. Teams can review retrieval outcomes, supported abstentions, latency, source coverage, permission failures, next-action selection, and replay instead of treating “the AI seemed helpful” as the acceptance criterion.

VALUE AND PROOFWhat leaders can measure, govern, and operationalize
  1. 01Faster discovery across products, policies, documentation, and institutional knowledge
  2. 02Lower onboarding and support cost through answer-first guided discovery
  3. 03Private search over sensitive material without exporting the complete corpus
  4. 04Provider independence because exact search and authority do not disappear with a model
  5. 05Quality gates for ranking, abstention, source coverage, latency, and route coherence
  6. 06Replayable evidence for audits, product improvement, and buyer confidence
ARCHITECTURE

Seven governed stages between a question and an answer you can defend.

Each stage has one job. Semantic systems broaden interpretation; exact software preserves eligibility, ordering, authority, and replay.

The architecture keeps proposal, evidence, policy, ranking, planning, and commitment separate. A useful embedding match cannot silently bypass rights, freshness, scope, official sources, or action permissions.

Authority boundary: similarity is not truth, a graph edge is not permission, a model response is not an approved action, and a renderer is never the system of record.

Canonical flowExact + semantic proposals → graph context → authority gate → deterministic fusion → Q-Lens → receipt and replay
  1. 01

    Exact / lexical proposals. Match literal terms, IDs, commands, source names, and known phrases immediately.

  2. 02

    MiniLM semantic proposals. Interpret concepts, intents, scenarios, entities, and outcomes without authorizing them.

  3. 03

    Reviewed graph expansion. Add connected concepts, comparisons, misconceptions, sources, and bounded next actions.

  4. 04

    OverLift Authority Gate. Enforce rights, currentness, support, scope, safety, and evidence eligibility.

  5. 05

    Deterministic fusion. Rank only admitted candidates with fixed weights, stable identities, and exact tie-breaks.

  6. 06

    Q-Lens. Choose a bounded answer, learning path, tool route, next action, or abstention from admitted evidence.

  7. 07

    Decision Receipt and replay. Preserve the ranking, exclusions, versions, action, and state transition for inspection.

ENGINEERING

Local workers, typed evidence, deterministic fallback, and inspectable ranking.

The search experience can improve semantically without blocking the exact lane, surrendering canonical state, or requiring a hosted model endpoint.

Same-origin workers isolate model preparation and inference from the interaction path. The exact lane remains usable while semantic assets load, retry, fail, or are deliberately disabled.

Reference stackASP.NET Core · JavaScript · Web Workers · MiniLM / ONNX · local vectors · reviewed graphs · C17/Wasm authority · IndexedDB / browser cache · receipts
IMPLEMENTED ENGINEERINGBounded components and reliability contracts proven in the reference products
  1. 01Deferred same-origin MiniLM initialization in an isolated worker
  2. 02Exact and lexical search remains available during prepare, retry, and fallback
  3. 03Versioned vector packs, stable candidate IDs, and source or corpus hashes
  4. 04Browser cache and IndexedDB persistence where the product permits it
  5. 05Raw-query non-persistence and minimum-disclosure external boundaries
  6. 06Deadlines, fresh-worker retry, cancellation, abstention, and deterministic replay
OVERLIFT METHOD

Semantic search becomes trustworthy when every stage has an authority boundary.

OverLift lets search become more adaptive without allowing model timing, cache state, corpus order, or provider output to redefine official truth.

  1. T0
    Exact truth and deoptimization target

    Literal search, rights, source identity, policy, canonical state, and fail-closed behavior remain exact.

  2. T1
    Ready lexical path

    Use the known exact route immediately when the wording, identifier, or source is unambiguous.

  3. T2
    Specialized semantic path

    Use a reviewed local semantic model to expand meaning over already bounded product evidence.

  4. T3
    Verified trace

    Join exact, semantic, and graph evidence into one inspectable, repeatable candidate trace.

  5. T4
    Guarded acceleration

    Move faster with workers, vectors, caches, and GPU-capable paths while retaining exact exits.

  6. T5
    Proven memory

    Reuse verified vector packs, reviewed winners, and quarantined shortcuts without preserving raw private questions.

  7. T6
    Meaning proof

    Turn exact candidate state into a supported explanation with sources, uncertainty, and citations.

  8. T7
    Causal proof

    Show not only what ranked, but which evidence, exclusions, rules, and state changes produced the result.

  9. Tier-S
    Shadow challenge and deoptimization

    Compare an experimental evaluator against production authority; reject, quarantine, or require approval before graduation.

CATS

Context-Aware Tier Selection

CATS chooses the least risky qualified search path for the query, evidence, rights, device, latency, model readiness, and failure state. The most sophisticated path does not run merely because it exists.

SEMANTIC BRIDGE

Human language becomes typed search state

The Semantic Bridge turns questions and observations into explicit intents, entities, evidence needs, uncertainty, source boundaries, tools, permissions, and legal candidate actions.

DETERMINISM

One admitted order, one official result, one receipt

Models may retrieve, classify, compare, or propose. Exact software owns eligibility, ranking policy, state transitions, tool approval, commitment, and replay.

Read the complete OverLift Method Explore the security architecture Run the OverLift demonstration Explore semantic search in the glossary
BROWSER-BASED AGENTIC AI SYSTEM

Search is the discovery fabric that gives agents evidence without giving them authority.

Semantic Search sits between human intent and governed action. It helps bounded agents find and organize evidence, while rights, policy, exact code, human approval, and deterministic receipts remain in control.

CAPABILITY PROOFHow discovery agents, tools, authority, and replay work together
  1. 01PROPOSAL

    Intent agent

    Interprets the question, likely objective, entities, constraints, and evidence needs without declaring the result.

  2. 02ACTIVE

    Retrieval agent

    Runs exact, lexical, vector, and source-bound RAG paths over the evidence the current user may access.

  3. 03BOUNDED

    Graph agent

    Expands reviewed relationships, comparisons, prerequisites, misconceptions, and possible next actions.

  4. 04AUTHORITY

    OverLift Authority Gate

    Rejects stale, unsupported, out-of-scope, unsafe, or permission-ineligible candidates before fusion or planning.

  5. 05PLANNING

    Q-Lens

    Chooses a bounded answer, route, question, tool proposal, or abstention from already admitted evidence.

  6. 06TYPED TOOL

    Tool and action boundary

    Search may propose a next operation, but typed permissions and human approval determine whether it can execute.

  7. 07T0 / WASM

    C17/WebAssembly exact kernel

    Provides a portable correctness target for admission, fusion, numerical policy, deoptimization, and replay-critical work.

  8. 08PROOF

    Decision Receipt and Replay Lab

    Preserves the question digest, versions, candidates, exclusions, ordering, chosen path, and exact committed outcome.

Evidence-bound RAG Source and identity checks Narrow agents Typed tools Sandboxed execution Explicit capability allowlists Least privilege Local-first privacy Minimum disclosure Human approval Abstention and fail-closed paths Receipts, replay, and evaluation
What this proves A semantic system can understand imperfect human language, discover connected evidence, preserve privacy, contain untrusted execution, support agents through explicit capabilities, and still keep official truth and consequential actions deterministic.
Discuss this system Read the definitive article Review Sean's resume See the reference products
PRODUCT PROOF

Two very different products prove the same governed search architecture.

Veil applies semantic search to intimate, interpretive questions. AzureGlossary applies it to technical learning, exact terminology, official sources, and certification decisions.

Veil product dashboard with celestial visualization, Ask Veil controls, and personal chart experience
Veil · domain semantic search for private personal questions
CASE STUDY · VEIL OD26E

Natural-language astrology questions become typed plans—not permission to rewrite the chart.

Ask Veil routes a person’s question through a typed planner, local vector evidence, semantic pathfinding, the Semantic Bridge, CATS and Q-Lens, then into a deterministic reading compiler. The latest OD26E release preserves that authority path while hardening storage, service-worker caching, Worker lifecycle, and browser/server security boundaries. Exact celestial positions and chart facts remain separate from authored interpretation.

  • 156 / 156human-answer families retained
  • 468 / 468question-universe cases
  • 420 / 420hard-language questions with zero unsafe acceptance
  • 288 documentsLocal Vector Fabric with 6 / 6 retrieval cases
  • Privacy-first local runtimezero external answer-LLM calls, bounded caches, storage admission, and Worker failure paths
Open the Veil product story
AzureGlossary command palette showing technical search and reviewed Azure command discovery
AzureGlossary · semantic discovery over exact technical knowledge
CASE STUDY · AZUREGLOSSARY DP900-FINAL-R6

Semantic relevance improves discovery while official Microsoft sources remain authority.

AzureGlossary combines exact and lexical proposals, browser-local MiniLM, reviewed graph expansion, an OverLift Authority Gate, deterministic fusion, Q-Lens, abstention, and receipts. Exact SQLite and lexical search remain usable while the semantic layer prepares or retries.

  • 885 recordsin the current AI-103 and DP-900 semantic corpus
  • 22.97 MB q8 MiniLMpinned model with browser cache and IndexedDB vector evidence
  • Raw queries not persisteda concrete privacy boundary in the search coordinator
  • 100% abstentionin the retained global owner-campaign evidence
  • 82 ms p95in the retained production-search campaign evidence
Open the AzureGlossary product story
WHAT THE TWO PROOFS ESTABLISHOne architecture, two very different authority domains
  1. 01Domain ontology and reviewed sources matter more than a generic model prompt
  2. 02Semantic search can remain local, private, and useful without making cloud inference mandatory
  3. 03Exact facts and authored interpretation can coexist without collapsing into one authority layer
  4. 04Abstention, exclusions, receipts, and replay are product features—not internal diagnostics