// llmindex.io — Prisma schema: models, eval items, responses, duels, score runs
// Author:  Simon-Pierre Boucher
// Contact: contact@spboucher.ai
// License: Proprietary — © Simon-Pierre Boucher, all rights reserved

generator client {
  provider = "prisma-client-js"
}

datasource db {
  provider = "postgresql"
  url      = env("DATABASE_URL")
}

// LLMs synced daily from OpenRouter /models (slug, pricing → cost_per_1k_items).
model Model {
  id              String   @id @default(cuid())
  slug            String   @unique // OpenRouter slug, e.g. "anthropic/claude-..."
  name            String
  provider        String
  contextLength   Int?     @map("context_length")
  // USD per 1M tokens (converted from OpenRouter per-token pricing at sync time)
  promptPricePerM Float?   @map("prompt_price_per_m")
  completionPricePerM Float? @map("completion_price_per_m")
  active          Boolean  @default(true)
  ranked          Boolean  @default(false) // curated subset shown on the index
  vision          Boolean  @default(false) // accepts image input (vision_ocr domain)
  syncedAt        DateTime @map("synced_at")
  createdAt       DateTime @default(now()) @map("created_at")

  responses ModelResponse[]
  scores    Score[]
  duelsAsA  PairwiseDuel[]  @relation("duel_model_a")
  duelsAsB  PairwiseDuel[]  @relation("duel_model_b")

  @@map("models")
}

// Instantiated evaluation items. answer_key/rubric are SERVER-SIDE ONLY:
// they must never be selected by web app queries or public API serializers.
model EvalItem {
  id          String   @id @default(cuid())
  templateId  String   @map("template_id") // versioned template id from packages/items
  domain      String
  prompt      String
  answerKey   String   @map("answer_key")
  rubric      String?
  perturbSeed String?  @map("perturb_seed") // seed used by the perturbation engine
  isAnchor    Boolean  @default(false) @map("is_anchor") // fixed longitudinal subset (≤20% of a run)
  irtA        Float?   @map("irt_a") // 2PL discrimination (from last fit)
  irtB        Float?   @map("irt_b") // 2PL difficulty (logits, from last fit)
  status      String   @default("active") // active | flagged_for_retirement | retired
  createdAt   DateTime @default(now()) @map("created_at")

  responses ModelResponse[]
  duels     PairwiseDuel[]

  @@index([domain, status])
  @@map("eval_items")
}

// One row per batch (kind=eval_batch|duel_batch) or per index fit (kind=index_fit).
// Historical runs are immutable; every displayed score traces to one of these.
model ScoreRun {
  id             String    @id @default(cuid())
  indexVersion   String    @map("index_version") // INDEX_VERSION from packages/scoring
  kind           String // eval_batch | duel_batch | index_fit | demo_seed
  status         String    @default("pending") // pending | running | complete | degraded | failed
  itemSetHash    String    @map("item_set_hash") // sha256 over ordered item ids + perturb seeds
  modelSet       Json      @map("model_set") // array of model slugs in the run
  fitDiagnostics Json?     @map("fit_diagnostics") // convergence, judge agreement, bias rates…
  notes          String?
  createdAt      DateTime  @default(now()) @map("created_at")
  completedAt    DateTime? @map("completed_at")

  responses ModelResponse[]
  duels     PairwiseDuel[]
  scores    Score[]

  @@index([kind, status, createdAt])
  @@map("score_runs")
}

// Full audit trail: a ranking without stored raw responses is invalid.
model ModelResponse {
  id              String   @id @default(cuid())
  runId           String   @map("run_id")
  modelId         String   @map("model_id")
  itemId          String   @map("item_id")
  sampleIndex     Int      @default(0) @map("sample_index") // >0 for consistency k-samples
  requestParams   Json     @map("request_params") // exact temperature, max_tokens, seed…
  rawResponse     Json     @map("raw_response")
  answerExtracted String?  @map("answer_extracted")
  correct         Boolean?
  confidence      Float? // model-reported confidence in [0,1] → calibration
  tokensIn        Int?     @map("tokens_in")
  tokensOut       Int?     @map("tokens_out")
  latencyMs       Int?     @map("latency_ms")
  costUsd         Float?   @map("cost_usd")
  error           String?
  createdAt       DateTime @default(now()) @map("created_at")

  run   ScoreRun @relation(fields: [runId], references: [id])
  model Model    @relation(fields: [modelId], references: [id])
  item  EvalItem @relation(fields: [itemId], references: [id])

  @@unique([runId, modelId, itemId, sampleIndex])
  @@index([modelId, createdAt])
  @@map("model_responses")
}

// LLM-judged pairwise duels feeding the Bradley-Terry layer.
model PairwiseDuel {
  id              String   @id @default(cuid())
  runId           String   @map("run_id")
  domain          String
  itemId          String   @map("item_id")
  modelAId        String   @map("model_a_id")
  modelBId        String   @map("model_b_id")
  judgeSlug       String   @map("judge_slug") // never a model judging its own duel
  positionSwapped Boolean  @default(false) @map("position_swapped")
  winner          String // "a" | "b" | "tie"
  rawJudgment     Json     @map("raw_judgment")
  createdAt       DateTime @default(now()) @map("created_at")

  run    ScoreRun @relation(fields: [runId], references: [id])
  item   EvalItem @relation(fields: [itemId], references: [id])
  modelA Model    @relation("duel_model_a", fields: [modelAId], references: [id])
  modelB Model    @relation("duel_model_b", fields: [modelBId], references: [id])

  @@index([runId, domain])
  @@map("pairwise_duels")
}

// Final scores. domain="global" for the Global Index; otherwise one of
// the domains in packages/scoring/src/domains.ts. Always with CI bounds.
model Score {
  id         String  @id @default(cuid())
  runId      String  @map("run_id")
  modelId    String  @map("model_id")
  domain     String // "global" | code | math | reasoning | …
  score      Float // 0–1000
  scoreLow   Float   @map("score_low")
  scoreHigh  Float   @map("score_high")
  // { accuracy_irt, consistency, calibration, contamination_delta, latency_p50, cost_per_1k_items }
  subMetrics Json?   @map("sub_metrics")

  run   ScoreRun @relation(fields: [runId], references: [id])
  model Model    @relation(fields: [modelId], references: [id])

  @@unique([runId, modelId, domain])
  @@index([domain])
  @@map("scores")
}