📋 Tender Assistant

Structured JSON rules — no eval(). Conditions use an allow-listed operator set (eq, ne, gt, gte, lt, lte, in, not_in, exists, not_exists, contains, regex_match, before_today, after_today, is_true, is_false); actions are typed (mark_status, add_warning, generate_clarification, set_field, block_until_human_review). Each rule has a version that bumps on body edits and a status (draft → active → archived); the evaluator only runs active rows.

A dry run of every active rule, in evaluation order. Testing rules one at a time cannot show that they interact — a new rule can shadow one that already fires. Nothing is saved.

Supply a JSON facts payload and we'll run the rule deterministically — no model call.

Registry of Tender AI agents (Discovery, Requirement Extraction, Compliance, Clarifications, Technical Proposal, OPIS, DUAE, Past Bid Reuse, Bid Reviewer, Anonymization, Executive Decision, Orchestrator, …). Each row carries its prompt-pack contract — JSON output schema, anonymization / evidence / human-review flags, default model + RAG + flow-step bindings. Runtime services resolve which agent to invoke per flow step from this table; missing wiring returns setup_required.

Replaces sensitive entities with stable numbered placeholders before any external LLM call. Detectors: EMAIL, PHONE, CUI, BANK_ACCOUNT (IBAN), CONTRACT_ID (keyword-anchored), plus list-driven COMPANY_NAME / CUSTOMER_NAME / PERSON_NAME / ADDRESS / PROJECT_REFERENCE. This preview is side-effect free — runtime calls record an audit event (counts only, never the raw map unless policy allows).

Tools exposed through the platform MCP layer. Safety invariants are baked in: can_submit_bid=false, can_sign_documents=false, can_certify_compliance=false, enforces_tenant_isolation=true on every tool. Tools that produce bid content carry human_review_required=true; factual tools carry requires_evidence_citation=true. The service layer refuses to persist a tool that lifts any of these.

Tender Assistant consumes tools from AI Admin, DMS, Workflow, Template Builder, Document Generator, Company Profile, RAG Service, OCR Service, Digital Signature, and Notification Service. The Digital Signature row is locked to data_privacy_policy="no_auto_sign" — the service layer refuses to add submit/auto-sign tools to its allow-list.

Tender services call this layer before invoking the AI Admin runtime. Routes record which model profile, fallback, RAG profile, MCP profiles, prompt version, and cost/privacy policies apply per (category, flow, step). The resolver picks the most specific match (cat+step → cat → step → global) and returns setup_required when no model profile is configured.

Reuses the platform-wide AI Admin training data table — Tender admin only owns the 6-group taxonomy and tags existing rows via the AI Admin meta JSON. Safety invariant: training_ready=true is only accepted when reviewer_status='approved' AND anonymization_status ∈ {anonymized, not_sensitive} — fine-tuning never sees raw sensitive content.

Reuses the platform-wide AI Admin Knowledge Base. Each group is a Tender-domain filter; KB rows are tagged into a group via the AI Admin KB tags JSON. The group taxonomy lives in the database and is editable here.

Ordered pipelines that drive Tender analysis. Runtime services resolve which flow + steps to execute by tenant + category. If no active flow exists for a category and no default flow is set, runtime returns setup_required — never silent hardcoded behavior.