Deploy: Stack A (NIM brain + Maverick judge + Sarvam voice). D-019.

40-data/llm_health.json

@@ -1,22 +1,16 @@
 {
-  "updated_at": "2026-05-15T01:52:30Z",
+  "updated_at": "2026-05-15T01:54:30Z",
   "models": {
     "qwen/qwen3-next-80b-a3b-instruct": {
       "model": "qwen/qwen3-next-80b-a3b-instruct",
       "status": "healthy",
-      "last_success_at": "2026-05-15T01:52:30Z",
+      "last_success_at": "2026-05-15T01:54:30Z",
       "last_failure_at": null,
       "last_error": null,
-      "latency_ms": 1208,
+      "latency_ms": 771,
       "consecutive_failures": 0,
-      "tested_at": "2026-05-15T01:52:30Z",
+      "tested_at": "2026-05-15T01:54:30Z",
       "probe_history": [
-        {
-          "ok": true,
-          "latency_ms": 458,
-          "ts": "2026-05-15T01:38:54Z",
-          "src": "probe"
-        },
         {
           "ok": true,
           "latency_ms": 285,
@@ -40,6 +34,12 @@
           "latency_ms": 1208,
           "ts": "2026-05-15T01:52:30Z",
           "src": "probe"
+        },
+        {
+          "ok": true,
+          "latency_ms": 771,
+          "ts": "2026-05-15T01:54:30Z",
+          "src": "probe"
         }
       ],
       "degraded_until_monotonic": 0.0,
@@ -53,18 +53,12 @@
       "model": "qwen/qwen3.5-122b-a10b",
       "status": "down",
       "last_success_at": null,
-      "last_failure_at": "2026-05-15T01:52:30Z",
+      "last_failure_at": "2026-05-15T01:54:30Z",
       "last_error": "timeout",
       "latency_ms": null,
-      "consecutive_failures": 17,
-      "tested_at": "2026-05-15T01:52:30Z",
+      "consecutive_failures": 18,
+      "tested_at": "2026-05-15T01:54:30Z",
       "probe_history": [
-        {
-          "ok": false,
-          "latency_ms": 8003,
-          "ts": "2026-05-15T01:38:54Z",
-          "src": "probe"
-        },
         {
           "ok": false,
           "latency_ms": 8001,
@@ -88,6 +82,12 @@
           "latency_ms": 8001,
           "ts": "2026-05-15T01:52:30Z",
           "src": "probe"
+        },
+        {
+          "ok": false,
+          "latency_ms": 8002,
+          "ts": "2026-05-15T01:54:30Z",
+          "src": "probe"
         }
       ],
       "degraded_until_monotonic": 0.0,
@@ -101,18 +101,12 @@
       "model": "openai/gpt-oss-120b",
       "status": "down",
       "last_success_at": null,
-      "last_failure_at": "2026-05-15T01:52:30Z",
+      "last_failure_at": "2026-05-15T01:54:30Z",
       "last_error": "empty_content",
       "latency_ms": null,
-      "consecutive_failures": 17,
-      "tested_at": "2026-05-15T01:52:30Z",
+      "consecutive_failures": 18,
+      "tested_at": "2026-05-15T01:54:30Z",
       "probe_history": [
-        {
-          "ok": false,
-          "latency_ms": 399,
-          "ts": "2026-05-15T01:38:54Z",
-          "src": "probe"
-        },
         {
           "ok": false,
           "latency_ms": 261,
@@ -136,6 +130,12 @@
           "latency_ms": 253,
           "ts": "2026-05-15T01:52:30Z",
           "src": "probe"
+        },
+        {
+          "ok": false,
+          "latency_ms": 311,
+          "ts": "2026-05-15T01:54:30Z",
+          "src": "probe"
         }
       ],
       "degraded_until_monotonic": 0.0,
@@ -147,20 +147,14 @@
     },
     "mistralai/mistral-large-3-675b-instruct-2512": {
       "model": "mistralai/mistral-large-3-675b-instruct-2512",
-      "status": "degraded",
-      "last_success_at": "2026-05-15T01:46:37Z",
+      "status": "healthy",
+      "last_success_at": "2026-05-15T01:54:30Z",
       "last_failure_at": "2026-05-15T01:52:30Z",
-      "last_error": "timeout",
-      "latency_ms": 449,
-      "consecutive_failures": 1,
-      "tested_at": "2026-05-15T01:52:30Z",
+      "last_error": null,
+      "latency_ms": 617,
+      "consecutive_failures": 0,
+      "tested_at": "2026-05-15T01:54:30Z",
       "probe_history": [
-        {
-          "ok": true,
-          "latency_ms": 709,
-          "ts": "2026-05-15T01:38:54Z",
-          "src": "probe"
-        },
         {
           "ok": true,
           "latency_ms": 782,
@@ -184,6 +178,12 @@
           "latency_ms": 8002,
           "ts": "2026-05-15T01:52:30Z",
           "src": "probe"
+        },
+        {
+          "ok": true,
+          "latency_ms": 617,
+          "ts": "2026-05-15T01:54:30Z",
+          "src": "probe"
         }
       ],
       "degraded_until_monotonic": 0.0,
@@ -195,20 +195,14 @@
     },
     "nvidia/llama-3.3-nemotron-super-49b-v1.5": {
       "model": "nvidia/llama-3.3-nemotron-super-49b-v1.5",
-      "status": "healthy",
-      "last_success_at": "2026-05-15T01:52:30Z",
+      "status": "degraded",
+      "last_success_at": "2026-05-15T01:54:30Z",
       "last_failure_at": "2026-05-15T01:24:45Z",
       "last_error": null,
-      "latency_ms": 349,
+      "latency_ms": 5838,
       "consecutive_failures": 0,
-      "tested_at": "2026-05-15T01:52:30Z",
+      "tested_at": "2026-05-15T01:54:30Z",
       "probe_history": [
-        {
-          "ok": true,
-          "latency_ms": 4820,
-          "ts": "2026-05-15T01:38:54Z",
-          "src": "probe"
-        },
         {
           "ok": true,
           "latency_ms": 385,
@@ -232,6 +226,12 @@
           "latency_ms": 349,
           "ts": "2026-05-15T01:52:30Z",
           "src": "probe"
+        },
+        {
+          "ok": true,
+          "latency_ms": 5838,
+          "ts": "2026-05-15T01:54:30Z",
+          "src": "probe"
         }
       ],
       "degraded_until_monotonic": 0.0,
@@ -244,19 +244,13 @@
     "meta/llama-3.3-70b-instruct": {
       "model": "meta/llama-3.3-70b-instruct",
       "status": "healthy",
-      "last_success_at": "2026-05-15T01:52:30Z",
+      "last_success_at": "2026-05-15T01:54:30Z",
       "last_failure_at": "2026-05-15T01:41:07Z",
       "last_error": null,
-      "latency_ms": 212,
+      "latency_ms": 3890,
       "consecutive_failures": 0,
-      "tested_at": "2026-05-15T01:52:30Z",
+      "tested_at": "2026-05-15T01:54:30Z",
       "probe_history": [
-        {
-          "ok": true,
-          "latency_ms": 511,
-          "ts": "2026-05-15T01:38:54Z",
-          "src": "probe"
-        },
         {
           "ok": false,
           "latency_ms": 8003,
@@ -280,6 +274,12 @@
           "latency_ms": 212,
           "ts": "2026-05-15T01:52:30Z",
           "src": "probe"
+        },
+        {
+          "ok": true,
+          "latency_ms": 3890,
+          "ts": "2026-05-15T01:54:30Z",
+          "src": "probe"
         }
       ],
       "degraded_until_monotonic": 0.0,
@@ -293,18 +293,12 @@
       "model": "deepseek-ai/deepseek-v4-pro",
       "status": "down",
       "last_success_at": null,
-      "last_failure_at": "2026-05-15T01:52:30Z",
+      "last_failure_at": "2026-05-15T01:54:30Z",
       "last_error": "timeout",
       "latency_ms": null,
-      "consecutive_failures": 17,
-      "tested_at": "2026-05-15T01:52:30Z",
+      "consecutive_failures": 18,
+      "tested_at": "2026-05-15T01:54:30Z",
       "probe_history": [
-        {
-          "ok": false,
-          "latency_ms": 8002,
-          "ts": "2026-05-15T01:38:54Z",
-          "src": "probe"
-        },
         {
           "ok": false,
           "latency_ms": 8059,
@@ -328,6 +322,12 @@
           "latency_ms": 8003,
           "ts": "2026-05-15T01:52:30Z",
           "src": "probe"
+        },
+        {
+          "ok": false,
+          "latency_ms": 8002,
+          "ts": "2026-05-15T01:54:30Z",
+          "src": "probe"
         }
       ],
       "degraded_until_monotonic": 0.0,
@@ -341,18 +341,12 @@
       "model": "nvidia/nemotron-3-nano-30b-a3b",
       "status": "down",
       "last_success_at": null,
-      "last_failure_at": "2026-05-15T01:52:30Z",
+      "last_failure_at": "2026-05-15T01:54:30Z",
       "last_error": "empty_content",
       "latency_ms": null,
-      "consecutive_failures": 17,
-      "tested_at": "2026-05-15T01:52:30Z",
+      "consecutive_failures": 18,
+      "tested_at": "2026-05-15T01:54:30Z",
       "probe_history": [
-        {
-          "ok": false,
-          "latency_ms": 267,
-          "ts": "2026-05-15T01:38:54Z",
-          "src": "probe"
-        },
         {
           "ok": false,
           "latency_ms": 271,
@@ -376,6 +370,12 @@
           "latency_ms": 263,
           "ts": "2026-05-15T01:52:30Z",
           "src": "probe"
+        },
+        {
+          "ok": false,
+          "latency_ms": 276,
+          "ts": "2026-05-15T01:54:30Z",
+          "src": "probe"
         }
       ],
       "degraded_until_monotonic": 0.0,
@@ -389,18 +389,12 @@
       "model": "deepseek-ai/deepseek-v4-flash",
       "status": "down",
       "last_success_at": null,
-      "last_failure_at": "2026-05-15T01:52:30Z",
+      "last_failure_at": "2026-05-15T01:54:30Z",
       "last_error": "timeout",
       "latency_ms": null,
-      "consecutive_failures": 17,
-      "tested_at": "2026-05-15T01:52:30Z",
+      "consecutive_failures": 18,
+      "tested_at": "2026-05-15T01:54:30Z",
       "probe_history": [
-        {
-          "ok": false,
-          "latency_ms": 8002,
-          "ts": "2026-05-15T01:38:54Z",
-          "src": "probe"
-        },
         {
           "ok": false,
           "latency_ms": 8002,
@@ -424,6 +418,12 @@
           "latency_ms": 8004,
           "ts": "2026-05-15T01:52:30Z",
           "src": "probe"
+        },
+        {
+          "ok": false,
+          "latency_ms": 8003,
+          "ts": "2026-05-15T01:54:30Z",
+          "src": "probe"
         }
       ],
       "degraded_until_monotonic": 0.0,
@@ -437,18 +437,12 @@
       "model": "moonshotai/kimi-k2-instruct-0905",
       "status": "down",
       "last_success_at": null,
-      "last_failure_at": "2026-05-15T01:52:30Z",
+      "last_failure_at": "2026-05-15T01:54:30Z",
       "last_error": "http_404",
       "latency_ms": null,
-      "consecutive_failures": 17,
-      "tested_at": "2026-05-15T01:52:30Z",
+      "consecutive_failures": 18,
+      "tested_at": "2026-05-15T01:54:30Z",
       "probe_history": [
-        {
-          "ok": false,
-          "latency_ms": 82,
-          "ts": "2026-05-15T01:38:54Z",
-          "src": "probe"
-        },
         {
           "ok": false,
           "latency_ms": 84,
@@ -472,6 +466,12 @@
           "latency_ms": 84,
           "ts": "2026-05-15T01:52:30Z",
           "src": "probe"
+        },
+        {
+          "ok": false,
+          "latency_ms": 81,
+          "ts": "2026-05-15T01:54:30Z",
+          "src": "probe"
         }
       ],
       "degraded_until_monotonic": 0.0,
@@ -485,18 +485,12 @@
       "model": "minimaxai/minimax-m2.5",
       "status": "down",
       "last_success_at": null,
-      "last_failure_at": "2026-05-15T01:52:30Z",
+      "last_failure_at": "2026-05-15T01:54:30Z",
       "last_error": "http_410",
       "latency_ms": null,
-      "consecutive_failures": 17,
-      "tested_at": "2026-05-15T01:52:30Z",
+      "consecutive_failures": 18,
+      "tested_at": "2026-05-15T01:54:30Z",
       "probe_history": [
-        {
-          "ok": false,
-          "latency_ms": 30,
-          "ts": "2026-05-15T01:38:54Z",
-          "src": "probe"
-        },
         {
           "ok": false,
           "latency_ms": 32,
@@ -520,6 +514,12 @@
           "latency_ms": 29,
           "ts": "2026-05-15T01:52:30Z",
           "src": "probe"
+        },
+        {
+          "ok": false,
+          "latency_ms": 31,
+          "ts": "2026-05-15T01:54:30Z",
+          "src": "probe"
         }
       ],
       "degraded_until_monotonic": 0.0,
@@ -532,19 +532,13 @@
     "meta/llama-4-maverick-17b-128e-instruct": {
       "model": "meta/llama-4-maverick-17b-128e-instruct",
       "status": "healthy",
-      "last_success_at": "2026-05-15T01:52:30Z",
+      "last_success_at": "2026-05-15T01:54:30Z",
       "last_failure_at": "2026-05-14T01:03:07Z",
       "last_error": null,
-      "latency_ms": 205,
+      "latency_ms": 247,
       "consecutive_failures": 0,
-      "tested_at": "2026-05-15T01:52:30Z",
+      "tested_at": "2026-05-15T01:54:30Z",
       "probe_history": [
-        {
-          "ok": true,
-          "latency_ms": 282,
-          "ts": "2026-05-15T01:38:54Z",
-          "src": "probe"
-        },
         {
           "ok": true,
           "latency_ms": 313,
@@ -568,6 +562,12 @@
           "latency_ms": 205,
           "ts": "2026-05-15T01:52:30Z",
           "src": "probe"
+        },
+        {
+          "ok": true,
+          "latency_ms": 247,
+          "ts": "2026-05-15T01:54:30Z",
+          "src": "probe"
         }
       ],
       "degraded_until_monotonic": 0.0,
@@ -581,18 +581,12 @@
       "model": "openrouter:openai/gpt-oss-120b",
       "status": "degraded",
       "last_success_at": "2026-05-15T01:46:37Z",
-      "last_failure_at": "2026-05-15T01:52:30Z",
+      "last_failure_at": "2026-05-15T01:54:30Z",
       "last_error": "empty_content",
       "latency_ms": 1349,
-      "consecutive_failures": 1,
-      "tested_at": "2026-05-15T01:52:30Z",
+      "consecutive_failures": 2,
+      "tested_at": "2026-05-15T01:54:30Z",
       "probe_history": [
-        {
-          "ok": false,
-          "latency_ms": 891,
-          "ts": "2026-05-15T01:38:54Z",
-          "src": "probe"
-        },
         {
           "ok": true,
           "latency_ms": 4642,
@@ -616,31 +610,31 @@
           "latency_ms": 1635,
           "ts": "2026-05-15T01:52:30Z",
           "src": "probe"
+        },
+        {
+          "ok": false,
+          "latency_ms": 856,
+          "ts": "2026-05-15T01:54:30Z",
+          "src": "probe"
         }
       ],
       "degraded_until_monotonic": 0.0,
       "credits_remaining": 0.0,
       "credits_unit": "usd_balance",
       "credits_reset_at": null,
-      "credits_observed_at": 5039.664027625,
+      "credits_observed_at": 5158.700559583,
       "credits_low_water": 0.05
     },
     "groq:llama-3.3-70b-versatile": {
       "model": "groq:llama-3.3-70b-versatile",
       "status": "healthy",
-      "last_success_at": "2026-05-15T01:52:30Z",
+      "last_success_at": "2026-05-15T01:54:30Z",
       "last_failure_at": null,
       "last_error": null,
-      "latency_ms": 247,
+      "latency_ms": 261,
       "consecutive_failures": 0,
-      "tested_at": "2026-05-15T01:52:30Z",
+      "tested_at": "2026-05-15T01:54:30Z",
       "probe_history": [
-        {
-          "ok": true,
-          "latency_ms": 154,
-          "ts": "2026-05-15T01:38:54Z",
-          "src": "probe"
-        },
         {
           "ok": true,
           "latency_ms": 285,
@@ -664,6 +658,12 @@
           "latency_ms": 247,
           "ts": "2026-05-15T01:52:30Z",
           "src": "probe"
+        },
+        {
+          "ok": true,
+          "latency_ms": 261,
+          "ts": "2026-05-15T01:54:30Z",
+          "src": "probe"
         }
       ],
       "degraded_until_monotonic": 0.0,

backend/main.py

@@ -1201,11 +1201,37 @@ async def policies_all(session_id: Optional[str] = None):
     # JSON that doesn't have an extracted counterpart yet. This way, even
     # policies whose LLM extraction failed still surface in the marketplace
     # with their human-curated data.
+    #
+    # KI-133 (2026-05-15) — dedup marketplace cards to ONE per policy product
+    # (not one per PDF document). Previously wordings + brochure + cis of
+    # the same product created three separate cards, ballooning the
+    # marketplace from 138 products to ~209 cards and confusing users who
+    # see "138 POLICIES" in the badge but 209 cards below.
+    #
+    # Algorithm: sort source filenames so the canonical "wordings" variant
+    # is encountered first for each product, then dedup by stripped
+    # policy_id (everything before the trailing __<doctype>).
+    # Doctype preference: wordings > prospectus > cis > brochure > anything
+    # else (alphabetical fallback).
+    _DOCTYPE_RANK = {"wordings": 0, "prospectus": 1, "cis": 2, "brochure": 3}
+    def _doctype_of(stem: str) -> str:
+        return stem.rsplit("__", 1)[1] if "__" in stem else ""
+    def _product_key_of(policy_id: str) -> str:
+        # Strip trailing __<doctype> so wordings/brochure/cis of the same
+        # product collapse to a single key.
+        return policy_id.rsplit("__", 1)[0] if "__" in policy_id else policy_id
+
+    sorted_files = sorted(
+        settings.EXTRACTED_DIR.glob("*.json"),
+        key=lambda fp: (_DOCTYPE_RANK.get(_doctype_of(fp.stem), 99), fp.stem),
+    )
+
+    seen_product_keys: set[str] = set()
     seen_policy_ids: set[str] = set()
     out = []
 
     # Pass 1: existing extracted policies (merged with curated overrides)
-    for fp in sorted(settings.EXTRACTED_DIR.glob("*.json")):
+    for fp in sorted_files:
         try:
             data = _json.loads(fp.read_text())
         except Exception:
@@ -1220,6 +1246,13 @@ async def policies_all(session_id: Optional[str] = None):
         # in chat answers, they just don't appear as marketplace cards.
         if slug == "regulatory":
             continue
+        # KI-133 (2026-05-15) — dedup by product (insurer__product), so the
+        # wordings PDF wins and the brochure/cis variants don't generate
+        # duplicate cards. Pass-1 sort order guarantees wordings comes first.
+        product_key = _product_key_of(policy_id_local)
+        if product_key in seen_product_keys:
+            continue
+        seen_product_keys.add(product_key)
         name, home = insurer_meta.get(slug, (slug, ""))
         # Get insurer reviews if available for the scorecard
         ir = None