Evidence Search Assistant — Technical reference

This reference owns the implementation contracts of BND-SEARCH. Read the package guide for usage, configuration, dependencies and upgrades. Generic execution, isolation and admission remain governed by Platform Documentation.

Data model and interfaces

DataModelPackage: DM-SEARCH

Owned entities: search session, question, search result reference, cited answer, saved question, monitor, history record.

InterfacePackage contributions: IF-SEARCH-ASSISTANT.

Each business object and source reference remains tenant- and purpose-scoped. The package owns schema evolution and declared migration behavior; the platform enforces schema registration, authorization and versioned installation. Consumers reference this model rather than creating competing canonical copies.

Workflows

WorkflowPackage: WFP-SEARCH-ASSISTANT

• Authorized Query Planning
• Hybrid Retrieval
• Cited Answer Generation
• Explicit Abstention
• Search Result Presentation
• Save Question
• Monitor Saved Question
• Search History Retention and Deletion

These are named definitions within the workflow package, not separate invented workflow IDs. Workflow runs resolve package versions, actor and purpose, authorized inputs, retries, review gates and output schemas through the environment lock. Temporal provides durable execution and Kafka carries declared lifecycle/domain events.

Policies, evaluations and tools

POL-SEARCH-DISCLOSURE, EVAL-RETRIEVAL-ANSWER, hybrid retrieval, citation, abstention, and saved-query tools.

Qualification covers the selected model/runtime profile, languages, access scope, failure handling and the package's domain-specific outcomes. Tool execution uses the platform gateway; package code cannot increase its own permissions or write directly to external systems.

Owned agents

U04 — Search Assistant Agent

Agent ID: effority.search-assistant

Purpose: Answers natural-language questions from authorized evidence with citations and explicit scope.

Triggers: User question, saved-question schedule, or application deep link.

Inputs: Authenticated actor, question, filters, authorized object scope, time range, language, and source permissions.

Outputs: Answer, citations, consulted scope, confidence, exclusions count, ambiguity request, or explicit NO_AUTHORIZED_EVIDENCE.

Tools: evidence.search, graph.read, timeline.read, document.read, citation.verify, and answer.render.

Review and limits: Source instructions are untrusted data. The agent does not reveal restricted-source existence and distinguishes no evidence from evidence that an event did not occur.

Primary metrics: Answer faithfulness, citation coverage, answer relevance, abstention correctness, permission violations, and user follow-up rate.

Implementation profiles

RAG Assistant Implementation

The question flow is:

1. authentication and ABAC decision;
2. language, intent, period, and object detection;
3. entity resolution with confirmation when ambiguous;
4. hybrid retrieval with prior ACL enforcement;
5. reranking and context assembly;
6. local answer generation with mandatory citations;
7. entailment and coverage verification;
8. final wording or abstention;
9. logging of the query, sources, and model according to policy.

If the verifier finds insufficient support, the assistant states that no evidence is available within the consulted scope. It distinguishes this result from asserting that the event never happened.

Documents are treated as untrusted data. Their instructions never enter the system prompt, the model receives no credentials, and every tool is mediated by services that recheck permissions and schema.

Embeddings, Indexes, and Hybrid Retrieval

The embedding service runs a registered local multilingual model with a normalized 1,024-dimensional output. The architecture can replace it without changing consumers; a migration builds a new index and retains the old one until validation completes.

Retrieval combines:

1. mandatory filters for tenant, ABAC, purpose, validity, object, and language;
2. BM25 for names, codes, figures, and lexical matches;
3. vector similarity for semantic equivalence;
4. Reciprocal Rank Fusion to combine candidates;
5. a local cross-encoder reranker over permitted candidates;
6. diversity to avoid ten redundant excerpts;
7. final selection by relevance, recency, source authority, and context budget.

Permission filtering happens before candidates are calculated or delivered to the model. A post-filter is never the sole access control. Citations preserve document, version, page or timestamp, and excerpt checksum.

The retrieval router also activates qualified profiles when the query and evidence require them:

• ColBERTv2-style late interaction for fine-grained transcript, clause, commitment, and acceptance retrieval;
• GraphRAG local, global, and DRIFT-style analysis over Telynt's versioned temporal business graph;
• RAPTOR-style hierarchical retrieval for long contracts, SOWs, reports, and meeting histories;
• ColPali-style visual retrieval for rendered proposals, slides, diagrams, tables, whiteboards, and screenshots;
• multilingual dense, sparse, and multi-vector embedding packages such as the qualified BGE-M3 profile.

These profiles supplement the hybrid core. Authorization is enforced before every candidate score and at every graph hop. Generated graph edges and hierarchical summaries remain derived artifacts; material claims resolve to original timestamped or paginated evidence.

Integration boundaries

Search does not grant access to sources. Saved questions and report consumers recheck permissions at execution and disclosure time.

Owned screen contracts: SCR-SEA-01, SCR-SEA-02, SCR-SEA-03, SCR-SEA-04.

See solution workflows for cross-package behavior. Screens, agents and workflows keep one canonical owner even when other bundles invoke or enrich them.