Jun 29 - Improve AI analysis function fix 429 error

routes/ai_summary.py

@@ -187,10 +187,6 @@ _OFFICE_ADDRESS = "2815 Hartland Road, Falls Church, VA 22043, USA"
 # JSON payload well under Groq's request size limit on all plan tiers.
 _TEXT_LIMIT = 14_000
 
-# Summary passed to stage 2. Stage 1 can produce up to ~16 KB of output;
-# capping it here prevents stage 2's payload from growing unbounded.
-_SUMMARY_LIMIT = 8_000
-
 # Analyst persona — injected as the system message in every API call.
 _SYSTEM_PROMPT = (
     "You are a senior government procurement analyst supporting a small business BD team. "
@@ -319,34 +315,25 @@ Our office: **{office}** — use as origin for travel estimates.
 DOCUMENTS:
 {documents}"""
 
-# Stage 2 — criteria evaluation (separate API call, only when active criteria exist).
+# Criteria evaluation — appended to the extraction prompt in the same API call.
-# Receives a capped slice of the stage 1 summary — not the raw documents — so the
+# A single call avoids the rate-limit (429) that two back-to-back calls trigger
-# model can focus on evaluation without re-parsing source noise.
+# on Groq's free tier.
-_CRITERIA_PROMPT = """\
+_CRITERIA_SUFFIX = """
-Below is an extracted summary of a government solicitation. Evaluate whether our
-company should pursue it based on the criteria listed.
-
-## OPPORTUNITY SUMMARY
-
-{summary}
 
 ---
 
-## ALIGNMENT EVALUATION
+## Opportunity Alignment Evaluation
 
-Our criteria:
+Now evaluate this opportunity against our company's criteria below.
+
+**Our criteria:**
 {criteria_list}
 
-For each criterion, state:
+For each criterion provide:
 - ✅ MEETS / ⚠️ PARTIALLY MEETS / ❌ DOES NOT MEET / ❓ CANNOT DETERMINE
-- One or two sentences citing specific details from the summary.
+- 1–2 sentences with specific evidence from the documents above.
 
----
+**Overall Recommendation** — write exactly one of these lines (machine-read, do not alter):
-
-## Overall Recommendation
-
-Write your recommendation on its own line in exactly one of these forms
-(machine-read — do not alter the label):
 
   RECOMMENDATION: PURSUE
   RECOMMENDATION: PASS
@@ -356,14 +343,26 @@ PURSUE = clearly aligns with most criteria and is competitive.
 PASS   = fails one or more critical criteria or presents unacceptable risk.
 UNCLEAR = insufficient information for a confident decision.
 
-## Executive Summary
+**Executive Summary:** 3–4 sentences in plain business language explaining the recommendation."""
-
-3–4 sentences in plain business language: strongest reasons to pursue or pass,
-and the single biggest risk or opportunity."""
 
 
-def _call_groq_api(api_key: str, model: str, user_message: str) -> str:
+def _call_groq(api_key: str, model: str, text: str, criteria: list) -> dict:
-    """Single Groq chat completions call. Raises on HTTP error."""
+    """Single Groq API call: extraction + optional criteria evaluation in one request."""
+    truncated = len(text) > _TEXT_LIMIT
+    doc_text  = text[:_TEXT_LIMIT]
+    n         = doc_text.count("=== ") or 1
+
+    prompt = _EXTRACTION_PROMPT.format(
+        n=n, office=_OFFICE_ADDRESS, documents=doc_text
+    )
+
+    if criteria:
+        criteria_list = "\n".join(
+            f"{i+1}. **{c['title']}**: {c['description']}"
+            for i, c in enumerate(criteria)
+        )
+        prompt += _CRITERIA_SUFFIX.format(criteria_list=criteria_list)
+
     response = http_requests.post(
         "https://api.groq.com/openai/v1/chat/completions",
         headers={
@@ -374,7 +373,7 @@ def _call_groq_api(api_key: str, model: str, user_message: str) -> str:
             "model":       model,
             "messages":    [
                 {"role": "system", "content": _SYSTEM_PROMPT},
-                {"role": "user",   "content": user_message},
+                {"role": "user",   "content": prompt},
             ],
             "max_tokens":  4096,
             "temperature": 0.1,
@@ -384,50 +383,19 @@ def _call_groq_api(api_key: str, model: str, user_message: str) -> str:
     if not response.ok:
         logger.error(f"Groq API {response.status_code}: {response.text[:300]}")
         response.raise_for_status()
-    return response.json()["choices"][0]["message"]["content"] or ""
 
-
+    content = response.json()["choices"][0]["message"]["content"] or ""
-def _call_groq(api_key: str, model: str, text: str, criteria: list) -> dict:
-    """
-    Two-stage Groq analysis:
-      Stage 1 (always):   extract structured fields + summary from the documents.
-      Stage 2 (optional): evaluate criteria against the capped stage-1 summary,
-                          not the raw docs, for a focused and reliable result.
-    """
-    truncated = len(text) > _TEXT_LIMIT
-    doc_text  = text[:_TEXT_LIMIT]
-    n         = doc_text.count("=== ") or 1
-
-    # ── Stage 1: extraction ────────────────────────────────────────────────────
-    stage1_prompt = _EXTRACTION_PROMPT.format(
-        n=n, office=_OFFICE_ADDRESS, documents=doc_text
-    )
-    summary = _call_groq_api(api_key, model, stage1_prompt)
-
-    if not criteria:
-        return {"verdict": None, "summary": summary, "truncated": truncated}
-
-    # ── Stage 2: criteria evaluation ───────────────────────────────────────────
-    criteria_list = "\n".join(
-        f"{i+1}. **{c['title']}**: {c['description']}"
-        for i, c in enumerate(criteria)
-    )
-    stage2_prompt = _CRITERIA_PROMPT.format(
-        summary=summary[:_SUMMARY_LIMIT],   # cap to control payload size
-        criteria_list=criteria_list,
-    )
-    evaluation = _call_groq_api(api_key, model, stage2_prompt)
 
     verdict = None
+    if criteria:
         match = re.search(
             r"RECOMMENDATION\s*:\s*(PURSUE|PASS|UNCLEAR)",
-        evaluation, re.IGNORECASE,
+            content, re.IGNORECASE,
         )
         if match:
             verdict = match.group(1).upper()
 
-    combined = summary + "\n\n---\n\n" + evaluation
+    return {"verdict": verdict, "summary": content, "truncated": truncated}
-    return {"verdict": verdict, "summary": combined, "truncated": truncated}
 
 
 # ─── Criteria Management (admin only) ─────────────────────────────────────────