Deploy v1 — single-Docker FastAPI + Next.js + RAG + voice + faithfulness

backend/providers/groq_llm.py

@@ -55,9 +55,25 @@ class GroqLLM(LLMProvider):
             "Content-Type": "application/json",
         }
 
+        import asyncio
+        # Groq free tier rate-limits aggressively (~30 req/min). Retry on 429
+        # with exponential backoff; also retry transient 5xx.
         async with httpx.AsyncClient(timeout=self.timeout) as client:
-            resp = await client.post(url, headers=headers, json=body)
-            resp.raise_for_status()
+            attempts = 4
+            delay = 1.5
+            for attempt in range(attempts):
+                resp = await client.post(url, headers=headers, json=body)
+                if resp.status_code == 429 or (500 <= resp.status_code < 600):
+                    if attempt == attempts - 1:
+                        resp.raise_for_status()
+                    # Honor Retry-After if Groq sends one; else exponential
+                    ra = resp.headers.get("Retry-After")
+                    wait = float(ra) if ra and ra.replace(".", "").isdigit() else delay
+                    await asyncio.sleep(wait)
+                    delay *= 2
+                    continue
+                resp.raise_for_status()
+                break
             payload = resp.json()
 
         choice = payload["choices"][0]

eval/run.py

@@ -125,13 +125,44 @@ Grade now."""
 
 
 async def run_one(gold: dict) -> EvalRecord:
-    turn = await handle_turn(
-        user_text=gold["question"],
-        chat_history=[],
-        user_profile={},
-        policy_filter_ids=[gold["policy_id"]],
-    )
-    factual, citation, score, reason = await grade_one(gold, turn.reply_text, turn.blocked)
+    """Single gold-question evaluation. Guarded so transient API errors (Groq
+    rate limit, network timeout) don't kill the whole sweep — the question
+    is recorded as failed and we move on."""
+    try:
+        turn = await handle_turn(
+            user_text=gold["question"],
+            chat_history=[],
+            user_profile={},
+            policy_filter_ids=[gold["policy_id"]],
+        )
+    except Exception as e:  # noqa: BLE001
+        msg = f"{type(e).__name__}: {str(e)[:200]}"
+        return EvalRecord(
+            id=gold["id"],
+            policy_id=gold["policy_id"],
+            question=gold["question"],
+            expected_answer=gold["expected_answer"],
+            bot_answer=f"[ORCHESTRATOR ERROR] {msg}",
+            factual_match=False,
+            citation_present=False,
+            judge_score=0.0,
+            judge_reason=f"orchestrator_error: {msg}",
+            expected_refusal=gold["expected_refusal"],
+            question_type=gold["question_type"],
+            difficulty=gold["difficulty"],
+            blocked=False,
+            faithfulness_passed=False,
+            faithfulness_reasons=[f"orchestrator_error: {msg}"],
+            brain_used="error",
+            latency_ms=0,
+        )
+    try:
+        factual, citation, score, reason = await grade_one(gold, turn.reply_text, turn.blocked)
+    except Exception as e:  # noqa: BLE001
+        factual = False
+        citation = bool(turn.citations) if hasattr(turn, "citations") else False
+        score = 0.0
+        reason = f"grader_error: {type(e).__name__}: {str(e)[:160]}"
     return EvalRecord(
         id=gold["id"],
         policy_id=gold["policy_id"],

tools/upload_to_hf.py

@@ -41,6 +41,13 @@ IGNORE = [
     "logs/**",
     "*.log",
     ".dockerignore",  # only needed at build time; HF doesn't use it
+    # Storage savings — HF Space free tier is 1GB. These regenerate at boot:
+    "rag/vectors/**",          # ~90MB Chroma DB; entrypoint.sh re-runs ingest
+    "rag/extracted/**",        # raw extraction transcripts; bot only reads .json
+    ".playwright-mcp/**",      # Browser snapshots from local Playwright sessions
+    "eval/results.json",       # Regenerated by eval runs
+    "eval/chunk_*.json",       # Sweep artifacts
+    "kb/calculations/chunk_*", # Sweep markdown
 ]