AI RFP Evaluation Agent: Build Guide

[RFP Document] → [Requirements Extractor] → [Scoring Rubric]
       ↓
[Vendor Submissions (PDF)] → [Compliance Scanner] → [Score per vendor]
       ↓
[Multi-Vendor Comparison Matrix] → [Decision Summary Report]

You are a government procurement evaluation expert.

Analyse this RFP document and produce a structured evaluation framework:

1. Mandatory Requirements (PASS/FAIL)
   - Label each M1, M2, M3... in order of appearance
   - Include the exact requirement text
   - Note any qualification threshold (e.g. "minimum 5 years experience")

2. Scored Requirements (scored out of stated maximum)
   - Label each S1, S2, S3...
   - Include the requirement, maximum points, and scoring guidance if stated

3. Red Flags to auto-detect
   - Items that would disqualify a submission regardless of score

4. Information gaps
   - Requirements that are ambiguous or missing scoring criteria

RFP Content:
{rfp_text}

Output as structured JSON.

You are evaluating a vendor's RFP submission for government procurement.

Evaluation Framework (from RFP):
{requirements_json}

Vendor Submission:
{vendor_submission_text}

For EACH mandatory requirement (M1, M2...):
- PASS: Clearly addressed with evidence
- FAIL: Not addressed or below threshold
- PARTIAL: Addressed but incomplete — note what is missing
- Quote the exact vendor text that addresses this requirement

For EACH scored requirement (S1, S2...):
- Score: X / [maximum points]
- Rationale: one sentence explaining the score
- Evidence: quote the vendor text supporting the score

Red Flags identified:
- List any disqualifying issues found

Overall readiness summary: 2-3 sentences on the vendor's submission quality.

Output as structured JSON.

You are producing a procurement evaluation matrix for committee review.

Vendor Evaluation Results:
{all_vendor_scores_json}

Evaluation Framework:
{requirements_json}

Produce:

1. Compliance Matrix
   - Rows: Mandatory requirements (M1, M2...)
   - Columns: Vendor names
   - Cells: PASS / FAIL / PARTIAL

2. Scoring Matrix
   - Rows: Scored requirements (S1, S2...)
   - Columns: Vendor names
   - Cells: Score achieved / Maximum
   - Bottom row: Total score per vendor

3. Red Flag Summary
   - Vendors with disqualifying issues

4. Ranking
   - Rank vendors by total score (excluding disqualified vendors)
   - Flag if scores are within 5% of each other (close call)

5. Decision Recommendation
   - Top-ranked vendor
   - Key differentiators from second-ranked vendor
   - Recommended next steps

Format as markdown tables. Keep language factual and audit-ready.

Review these vendor evaluation results for process integrity issues:

{all_vendor_scores_json}

Check for:
1. Any vendor with FAIL on a mandatory requirement (disqualified — confirm)
2. Scoring outliers (one vendor scored significantly higher — flag for review)
3. Incomplete submissions (vendor did not address >20% of requirements)
4. Missing information preventing fair comparison

Produce a one-page pre-committee briefing noting these issues.

Generate a formal Procurement Evaluation Report suitable for committee signature.

Format:
- Tender reference and title
- Evaluation date and panel members
- Summary of submissions received
- Methodology note (AI-assisted initial evaluation, human review)
- Compliance matrix (table)
- Scoring summary (table)
- Disqualified vendors and reasons
- Ranked shortlist with scores
- Recommended award and justification
- Dissenting notes field (leave blank for human completion)
- Sign-off block

Evaluation Data:
{final_evaluation_json}

Tone: formal, neutral, audit-ready.

import anthropic
from pypdf import PdfReader

client = anthropic.Anthropic()

def extract_text(pdf_path):
    reader = PdfReader(pdf_path)
    return "\n".join(page.extract_text() for page in reader.pages)

def evaluate_vendor(rfp_requirements, vendor_pdf):
    vendor_text = extract_text(vendor_pdf)
    response = client.messages.create(
        model="claude-opus-4-8",
        max_tokens=4096,
        messages=[{"role": "user", "content": EVAL_PROMPT.format(
            requirements_json=rfp_requirements,
            vendor_submission_text=vendor_text
        )}]
    )
    return response.content[0].text