Jun 29 - Improve AI analysis function

routes/ai_summary.py

@@ -182,97 +182,199 @@ def _extract_text(file_obj, ext: str) -> str:
 # Office address used as origin for distance/travel-time estimates.
 _OFFICE_ADDRESS = "2815 Hartland Road, Falls Church, VA 22043, USA"
 
+# Text character limit sent to the API. llama-3.3-70b supports 128K tokens (~500K chars);
+# 60K chars is a safe ceiling that leaves room for the prompt and response.
+_TEXT_LIMIT = 60_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. "
+    "Your role is to extract, organize, and evaluate information from solicitation documents "
+    "with precision and clarity. Format all output in well-structured Markdown using tables "
+    "and headers. Never invent or assume information not present in the source documents — "
+    "write \"Not specified\" for any missing field."
+)
+
 # Stage 1 — extraction prompt (always sent)
-_EXTRACTION_PROMPT = """You are an expert government procurement analyst.
+_EXTRACTION_PROMPT = """\
-The user has provided {n} document(s). Treat all provided content as a single combined source — do not analyze each file individually. Extract the following information once, consolidating details from all documents.
+Analyze the {n} attached government solicitation document(s) as one combined package.
+Complete every section below exactly as structured. Write **"Not specified"** for any field
+not found in the documents. Do not guess or invent details.
 
-IMPORTANT — Our office is located at:
+Our office is located at: **{office}** — use this as the origin for all travel estimates.
-  {office}
-Use this as the ORIGIN address for all driving distance and travel time calculations in field #9 below.
 
-Extract and clearly label the following fields (write "N/A" if not found):
+---
 
-  1. Solicitation Number
+## Solicitation Overview
-  2. Solicitation Type  (e.g. RFP, RFQ, IFB, etc.)
-  3. Set-Aside          (e.g. Small Business, 8(a), N/A)
-  4. Description / Scope of Work
-  5. Work Site / Location(s)
-  6. Pre-Proposal Conference / Site-Visit  (date, time, full address)
-  7. Point of Contact (POC)  (name, phone, email)
-  8. Total Square Footage    (if applicable)
-  9. Driving Distance & Travel Time
-     - Origin: {office}
-     - Destination: Pre-Proposal Conference or primary Work Site address
-     - Provide your best estimate of driving distance (miles) and typical driving time using major highways
-     - Note that these are AI estimates; actual times may vary with traffic
- 10. Last Day to Submit Questions
- 11. Due Date & Time
- 12. Any other notable requirements or deadlines
 
-Then provide a detailed OVERALL SUMMARY covering:
+| Field | Value |
+|---|---|
+| **Solicitation Number** | |
+| **Solicitation Type** | (RFP / RFQ / IFB / IDIQ / BPA / etc.) |
+| **Issuing Agency / Office** | |
+| **Set-Aside** | (Small Business / 8(a) / SDVOSB / HUBZone / WOSB / Unrestricted / etc.) |
+| **Contract Type** | (Firm-Fixed-Price / T&M / Cost-Plus / IDIQ / etc.) |
+| **Contract Period** | |
+| **NAICS Code** | |
+| **Estimated Value** | |
 
-  A. Scope of Work
+---
-  B. Contract Period
-  C. Proposal Submission Requirements
-  D. Key Deadlines & Action Items
 
-Be precise, detailed, and use bullet points throughout.
+## Scope of Work
-If information is not explicitly stated in the documents, note it as "Not specified in the document."
+
+*Describe what is required in full detail: services, deliverables, performance standards, and any technical requirements. Use bullet points.*
+
+---
+
+## Performance Location(s)
+
+*List all work sites. Note whether remote or on-site work is permitted.*
+
+---
+
+## Key Dates & Deadlines
+
+| Milestone | Date & Time (with timezone) |
+|---|---|
+| Pre-Proposal Conference / Site Visit | |
+| Questions Due | |
+| **Proposal Due** | |
+| Award Date (if stated) | |
+| Period of Performance Start | |
+| Other Deadlines | |
+
+---
+
+## Pre-Proposal Conference / Site Visit
+
+| | |
+|---|---|
+| **Date & Time** | |
+| **Full Address** | |
+| **Attendance** | (Mandatory / Optional / Not applicable) |
+| **RSVP Required** | (Yes — deadline & method / No) |
+
+---
+
+## Point of Contact (POC)
+
+| | |
+|---|---|
+| **Name / Title** | |
+| **Phone** | |
+| **Email** | |
+| **Questions Submission Method** | |
+
+---
+
+## Proposal Requirements
+
+*Summarize format, page limits, required sections/volumes, submission method (portal/email/mail), and number of copies. Use bullet points.*
+
+---
+
+## Evaluation Factors
+
+*List evaluation criteria and their weights or order of priority as stated in the solicitation.*
+
+---
+
+## Travel & Logistics
+
+| | |
+|---|---|
+| **Origin** | {office} |
+| **Destination** | (Pre-Proposal / Primary Work Site address) |
+| **Estimated Driving Distance** | |
+| **Estimated Drive Time** | (normal traffic conditions) |
+
+*(AI estimate — verify with a mapping service before scheduling travel.)*
+
+---
+
+## Notable Requirements & Red Flags
+
+*List anything that affects bid/no-bid decisions: unusual insurance or bonding levels, required security clearances, certifications, teaming or subcontracting restrictions, incumbent advantage indicators, aggressive timelines, or any other risk factors.*
+
+---
+
+## Total Square Footage
+
+*(If applicable to the scope of work; write "N/A" if not.)*
+
+---
+
+## Overall Summary
+
+Provide a concise but thorough summary covering:
+
+- **What is Being Procured** — Scope and key deliverables.
+- **Contract Period & Estimated Value** — Duration and any stated ceiling or estimate.
+- **Key Constraints & Requirements** — Timeline, location, special certifications, etc.
+- **Immediate Action Items** — What the team must do and by when.
+
+---
 
 DOCUMENTS:
 {documents}"""
 
-# Stage 2 — criteria evaluation suffix (appended only when active criteria exist)
+# Stage 2 — criteria evaluation (sent as a separate API call when active criteria exist).
-_CRITERIA_PROMPT_SUFFIX = """
+# Receives the clean Stage 1 extraction output as {summary}, not the raw documents,
+# so the model can focus entirely on the evaluation without re-parsing document noise.
+_CRITERIA_PROMPT = """\
+Below is an extracted and summarized government solicitation opportunity.
+Evaluate whether our company should pursue it based on our evaluation criteria.
 
-================================================================================
+---
-OPPORTUNITY ALIGNMENT EVALUATION
-================================================================================
 
-After completing the extraction and summary above, evaluate whether this
+## OPPORTUNITY SUMMARY
-opportunity aligns with our company's interests based on the following criteria.
+
+{summary}
+
+---
+
+## ALIGNMENT EVALUATION
+
+**Our evaluation criteria:**
 
-OUR EVALUATION CRITERIA:
 {criteria_list}
 
-For EACH criterion above:
+For **each criterion** listed above, provide:
-  - State whether the opportunity MEETS, DOES NOT MEET, or PARTIALLY MEETS it.
+- A status: ✅ **MEETS** / ⚠️ **PARTIALLY MEETS** / ❌ **DOES NOT MEET** / ❓ **CANNOT DETERMINE**
-  - Provide a brief, specific explanation citing details from the document(s).
+- A 1–2 sentence explanation citing specific details from the summary above.
 
-Then provide an OVERALL RECOMMENDATION using EXACTLY one of these three labels
+---
-on its own line (this label is machine-read — do not alter it):
 
-  RECOMMENDATION: PURSUE
+## Overall Recommendation
-  RECOMMENDATION: PASS
-  RECOMMENDATION: UNCLEAR
 
-Use PURSUE if the opportunity clearly meets most criteria and presents strong
+State your recommendation on its own line in **exactly** this format (required for system parsing — do not alter the label):
-alignment. Use PASS if it clearly fails key criteria. Use UNCLEAR if the
-documents lack sufficient information to make a confident determination.
 
-End with a 2-3 sentence EXECUTIVE SUMMARY explaining your recommendation
+```
-in plain business language."""
+RECOMMENDATION: PURSUE
+```
+or
+```
+RECOMMENDATION: PASS
+```
+or
+```
+RECOMMENDATION: UNCLEAR
+```
+
+**PURSUE** — opportunity clearly aligns with most criteria and is competitive.
+**PASS** — fails one or more critical criteria or presents unacceptable risk.
+**UNCLEAR** — insufficient information to make a confident determination.
+
+---
+
+## Executive Summary
+
+Write 3–4 sentences in plain business language explaining your recommendation:
+the strongest reasons to pursue or pass, and the single biggest risk or opportunity."""
 
 
-def _call_groq(api_key: str, model: str, text: str, criteria: list) -> dict:
+def _call_groq_api(api_key: str, model: str, user_message: str) -> str:
-    """Call the Groq chat completions REST API directly (no SDK required)."""
+    """Single Groq chat completions call. Raises on HTTP error."""
-    import re
-
-    n = text.count("=== ") or 1  # count file separators for the prompt header
-    truncated = len(text) > 14000
-
-    # Build the two-stage prompt matching the desktop app exactly
-    prompt = _EXTRACTION_PROMPT.format(
-        n=n, office=_OFFICE_ADDRESS, documents=text[:14000]
-    )
-
-    if criteria:
-        criteria_list = "\n".join(
-            f"  {i+1}. {c['title']}: {c['description']}"
-            for i, c in enumerate(criteria)
-        )
-        prompt += _CRITERIA_PROMPT_SUFFIX.format(criteria_list=criteria_list)
-
     response = http_requests.post(
         "https://api.groq.com/openai/v1/chat/completions",
         headers={
@@ -281,29 +383,61 @@ def _call_groq(api_key: str, model: str, text: str, criteria: list) -> dict:
         },
         json={
             "model":       model,
-            "messages":    [{"role": "user", "content": prompt}],
+            "messages":    [
-            "max_tokens":  4096,
+                {"role": "system", "content": _SYSTEM_PROMPT},
-            "temperature": 0.2,
+                {"role": "user",   "content": user_message},
+            ],
+            "max_tokens":  8192,
+            "temperature": 0.1,
         },
-        timeout=90,
+        timeout=120,
     )
     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 ""
 
-    # Parse the machine-readable RECOMMENDATION label (only present when criteria used)
+def _call_groq(api_key: str, model: str, text: str, criteria: list) -> dict:
+    """
+    Two-stage Groq analysis:
+      Stage 1 (always):   extract structured fields + overall summary from documents.
+      Stage 2 (optional): evaluate criteria against the Stage 1 output (not raw docs),
+                          producing a focused RECOMMENDATION with per-criterion scoring.
+    """
+    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 (uses clean summary, not raw docs) ────────
+    criteria_list = "\n".join(
+        f"{i+1}. **{c['title']}**: {c['description']}"
+        for i, c in enumerate(criteria)
+    )
+    stage2_prompt = _CRITERIA_PROMPT.format(
+        summary=summary, 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)",
-            content, re.IGNORECASE,
+        evaluation, re.IGNORECASE,
     )
     if match:
         verdict = match.group(1).upper()
 
-    return {"verdict": verdict, "summary": content, "truncated": truncated}
+    combined = summary + "\n\n---\n\n" + evaluation
+    return {"verdict": verdict, "summary": combined, "truncated": truncated}
 
 
 # ─── Criteria Management (admin only) ─────────────────────────────────────────