vc-finder

VC Finder

输入产品 URL 或描述。识别行业和阶段。查找 5 家获得融资的可比公司。开展两条研究线：谁投资了这些可比公司（Track A），以及哪些 VC 发布了关于这一领域的投资论点（Track B）。返回一份有来源、经过排序的投资者名单，并附上联系切入点。

---

零幻觉政策： 输出中的每条事实都必须能追溯到具体的 Tavily 搜索结果或抓取的产品页面。适用于：

• 可比公司名称：必须出现在 Tavily 搜索结果中，而非来自 AI 训练知识
• VC 基金名称：必须原样出现在 Tavily 搜索结果中
• 投资金额、阶段偏好、投资组合公司：必须来自搜索摘要，而非 AI 知识
• 基金概览和投资论点摘要：只能从搜索摘要中提取。如果搜索数据中没有某项细节，请写“搜索数据中未找到”，不要用训练知识补充。

---

常见错误

Agent 会想要……	为什么这是错的
因为 a16z 或 Sequoia 很有名就把它们加进去	没有证据的知名 VC 只会带来噪音。只纳入针对这一特定产品的 Tavily 搜索结果中出现的 VC。随意提及名号只会浪费创始人的时间。
根据训练知识生成可比公司	可比公司必须来自 Tavily 搜索结果（步骤 6）。AI 对公司的了解并非证据——凭记忆推荐的公司可能融资状态有误，或根本不是真正的可比公司。
当 Track A 的 5 次搜索全部返回 0 条结果时继续	Track A 结果为零，意味着可比公司选错了或过于冷门。请停止，使用更宽泛的搜索查询重新执行步骤 6，然后重试。
在未引用文章或帖子的情况下加入 Track B 的 VC	没有来源的投资论点与幻觉无法区分。创始人无法核实，名单也会失去全部可信度。
根据训练知识补充基金概览	基金概览必须仅来自 Tavily 摘要文本。如果摘要没有描述该基金，请写“搜索数据中未找到”。
根据网站的视觉风格判断阶段	阶段必须根据步骤 4 中识别出的具体 CTA 信号来判断。
编写泛泛的联系切入点	每个联系切入点都必须提及该产品的具体差异化优势，以及搜索数据中的具体 VC 投资组合信号或投资论点引述。
用户同时提供描述时跳过 URL 抓取	始终抓取 URL。实时页面通常会揭示用户描述中遗漏的阶段信号。

---

步骤 1：设置检查

echo "TAVILY_API_KEY:    ${TAVILY_API_KEY:+set}"
echo "FIRECRAWL_API_KEY: ${FIRECRAWL_API_KEY:-not set, Tavily extract will be used as fallback}"

如果缺少 TAVILY_API_KEY： 停止。告知用户：“研究 VC 投资和投资论点需要 TAVILY_API_KEY。没有备用方案。请在 app.tavily.com 获取——免费层级：每月 1000 个额度（约可完整运行 125 次）。将其添加到你的 .env 文件中。”

如果仅缺少 FIRECRAWL_API_KEY： 静默继续。URL 抓取将使用 Tavily extract。

---

步骤 2：收集输入

你需要：

• 产品 URL（必需，除非用户直接粘贴产品描述）
• 可选：目标阶段提示（pre-seed、seed、series-a、series-b）——如果提供，则使用该阶段并跳过阶段检测
• 可选：地域偏好（US、Europe、global）——未指定时默认为 US

如果用户只提供粘贴的描述（没有 URL）： 跳过步骤 3-4。直接进入步骤 5，将粘贴的文本作为 product_content。将 stage_source 设为 user_description。

如果既未提供 URL，也未提供描述： 询问：“你的产品或初创公司的 URL 是什么？或者粘贴一段简短描述：它的功能、面向的用户，以及你所处的阶段（pre-seed、seed、Series A）。”

从 URL 中提取产品 slug，用作输出文件名：

PRODUCT_SLUG=$(python3 -c "
from urllib.parse import urlparse
url = 'URL_HERE'
host = urlparse(url).netloc.replace('www.', '')
print(host.split('.')[0])
")

---

步骤 3：获取产品页面

首选：Firecrawl（如果已设置 FIRECRAWL_API_KEY）

curl -s -X POST https://api.firecrawl.dev/v1/scrape \
  -H "Authorization: Bearer $FIRECRAWL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url": "URL_HERE", "formats": ["markdown"], "onlyMainContent": true}' \
  | python3 -c "
import sys, json
d = json.load(sys.stdin)
content = d.get('data', {}).get('markdown', '') or d.get('markdown', '')
print(f'Fetched: {len(content)} characters')
open('/tmp/vc-product-raw.md', 'w').write(content)
"

备用：Tavily extract（如果未设置 FIRECRAWL_API_KEY）

curl -s -X POST https://api.tavily.com/extract \
  -H "Content-Type: application/json" \
  -d "{\"api_key\": \"$TAVILY_API_KEY\", \"urls\": [\"URL_HERE\"]}" \
  | python3 -c "
import sys, json
d = json.load(sys.stdin)
content = d.get('results', [{}])[0].get('raw_content', '')
print(f'Fetched via Tavily extract: {len(content)} characters')
open('/tmp/vc-product-raw.md', 'w').write(content)
"

步骤级检查点：

python3 -c "
content = open('/tmp/vc-product-raw.md').read()
if len(content) < 200:
    print('ERROR: Page returned fewer than 200 characters.')
else:
    print(f'Content OK: {len(content)} characters')
"

**如果内容少于 200 个字符：**停止获取。告知用户：“产品页面未返回可读内容。这通常意味着网站由 JavaScript 渲染，需要使用浏览器。请直接粘贴产品描述：它的功能、目标用户，以及你目前所处的阶段。”

使用粘贴的描述作为 product_content，继续执行步骤 5。

---

步骤 4：在本地检测阶段信号（无需 API）

在分析步骤之前，使用正则表达式解析获取到的 Markdown。

python3 << 'PYEOF'
import re, json

content = open('/tmp/vc-product-raw.md').read().lower()
stage_signals = []

if re.search(r'join\s+(the\s+)?waitlist|sign\s+up\s+for\s+beta|early\s+access|request\s+(an?\s+)?invite|get\s+notified', content):
    stage_signals.append({'signal': 'waitlist or beta CTA', 'stage_hint': 'pre-seed'})

if re.search(r'start\s+(your\s+)?free\s+trial|try\s+(it\s+)?for\s+free|request\s+a?\s+demo|book\s+a?\s+demo|schedule\s+a?\s+demo', content):
    stage_signals.append({'signal': 'free trial or demo CTA', 'stage_hint': 'seed'})

if re.search(r'contact\s+sales|talk\s+to\s+(our\s+)?sales|see\s+pricing|view\s+pricing|plans\s+and\s+pricing', content):
    stage_signals.append({'signal': 'pricing or sales CTA', 'stage_hint': 'series-a'})
if re.search(r'case\s+stud(y|ies)|customer\s+stor(y|ies)|trusted\s+by\s+[\d,]+|used\s+by\s+[\d,]+', content):
    stage_signals.append({'signal': 'case studies or customer count', 'stage_hint': 'series-a'})

if re.search(r'enterprise\s+(plan|pricing|tier)|we.?re\s+hiring|join\s+our\s+team|open\s+positions', content):
    stage_signals.append({'signal': 'enterprise tier or job openings', 'stage_hint': 'series-a-or-b'})

funding_match = re.search(
    r'raised\s+\$[\d,.]+\s*[mk]?|series\s+[abc]\s+round|seed\s+round|(\$[\d,.]+\s*[mk]?\s+(?:seed|series\s+[abc]))',
    content
)
if funding_match:
    stage_signals.append({'signal': f'funding text: {funding_match.group(0).strip()}', 'stage_hint': 'announced'})

if not stage_signals:
    dominant = 'unknown'
elif any(s['stage_hint'] == 'announced' for s in stage_signals):
    dominant = 'announced'
elif any(s['stage_hint'] == 'series-a-or-b' for s in stage_signals):
    dominant = 'series-a'
elif any(s['stage_hint'] == 'series-a' for s in stage_signals):
    dominant = 'series-a'
elif any(s['stage_hint'] == 'seed' for s in stage_signals):
    dominant = 'seed'
else:
    dominant = 'pre-seed'

confidence = 'high' if len(stage_signals) >= 2 else ('medium' if len(stage_signals) == 1 else 'low')

result = {'signals': stage_signals, 'dominant_stage': dominant, 'confidence': confidence}
json.dump(result, open('/tmp/vc-stage-signals.json', 'w'), indent=2)
print(f'Stage: {dominant} ({confidence} confidence) from {len(stage_signals)} signal(s)')
for s in stage_signals:
    print(f'  - {s["signal"]} -> {s["stage_hint"]}')
PYEOF

---

步骤 5：产品分析（分类体系、阶段、ICP）

输出产品内容和阶段信号：

python3 -c "
import json
content = open('/tmp/vc-product-raw.md').read()[:6000]
signals = json.load(open('/tmp/vc-stage-signals.json'))
print('=== PRODUCT PAGE (first 6000 chars) ===')
print(content)
print()
print('=== DETECTED STAGE SIGNALS ===')
print(json.dumps(signals, indent=2))
"

**AI 指令：**分析上方的产品页面内容。仅生成分类体系、ICP 和阶段分类——暂不生成可比公司（将在步骤 6 中通过实时搜索完成）。

写入 /tmp/vc-product-analysis.json：

• product_name：来自页面
• one_line_description：说明产品的功能、目标用户和核心价值主张。少于 20 个词。不要使用营销话术。
• industry_taxonomy：l1（顶级类别：金融科技 / 医疗科技 / 开发者工具 / 消费者等）、l2（行业类别：销售技术 / 物流软件等）、l3（具体细分领域：主动拓客 / 最后一公里路线规划等）。仅使用“技术”或“软件”等模糊标签不可接受。
• icp：buyer_persona（职位）、company_type、company_size
• detected_stage：pre-seed / seed / series-a / series-b / unknown
• stage_confidence：high / medium / low
• stage_evidence：用一句话准确指出页面上促成此判断的 CTA 或文本。如果未知，请填写“no clear signals found”。
• geography_bias：US / Europe / global / unclear
• comparable_companies：留空数组 []——将在步骤 6 中填入

python3 << 'PYEOF'
import json

analysis = {
    # FILL from your analysis above
    "comparable_companies": []
}

json.dump(analysis, open('/tmp/vc-product-analysis.json', 'w'), indent=2)
print('Product analysis written.')
PYEOF

验证：

python3 -c "
import json
a = json.load(open('/tmp/vc-product-analysis.json'))
print('Product:', a['product_name'])
print('Industry:', a['industry_taxonomy']['l1'], '>', a['industry_taxonomy']['l2'], '>', a['industry_taxonomy']['l3'])
print('Stage:', a['detected_stage'], '(' + a['stage_confidence'] + ' confidence)')
"

---

步骤 5b：与经过验证的基金数据集进行精选预匹配

将产品分类体系与包含 25 家经过验证的 VC 基金（来源于基金网站）的精选数据集进行比对。为 Track A 生成零幻觉的基金匹配和种子轮可比公司——不消耗 Tavily 点数。

输出产品分析结果以映射标签：

python3 -c "
import json
a = json.load(open('/tmp/vc-product-analysis.json'))
print('Taxonomy:', a['industry_taxonomy']['l1'], '>', a['industry_taxonomy']['l2'], '>', a['industry_taxonomy']['l3'])
print('Stage:', a['detected_stage'])
print('Geography:', a['geography_bias'])
"

**AI 指令：**将产品分类体系映射到基金数据集使用的标准标签。可用标签： DevTools、Infrastructure、Open Source、B2B SaaS、AI、Data、FinTech、HealthTech、Enterprise、Consumer、Marketplaces、E-commerce、Crypto、DeepTech、Cybersecurity、Generalist

选择 2-4 个能够描述该产品的标签。将 detected_stage 映射为：Pre-seed、Seed、Series A 或 Growth。将 geography_bias 映射为：US、Europe、India 或 Global。

写入产品上下文：

python3 << 'PYEOF'
import json

# FILL based on taxonomy analysis above
context = {
    "extracted_tags": ["TagA", "TagB"],  # 2-4 tags from the list above
    "stage_hint": "Seed",               # Pre-seed / Seed / Series A / Growth
    "geography_hint": "US"              # US / Europe / India / Global
}
json.dump(context, open('/tmp/vc-product-context.json', 'w'), indent=2)
print('Product context:', context)
PYEOF

使用嵌入的精选数据集进行评分：

python3 << 'PYEOF'
import json

context = json.load(open('/tmp/vc-product-context.json'))

VC_FUNDS = [
  {"fund_name":"Y Combinator","thesis":"We provide seed funding for startups. We invest in deeply technical teams building massive companies across all domains.","check_size":"$500k","stage_focus":["Pre-seed","Seed"],"industry_tags":["Generalist","B2B SaaS","DevTools","AI"],"geography_focus":["Global"],"notable_portfolio":["Stripe","Airbnb","GitLab"],"website":"https://www.ycombinator.com"},
  {"fund_name":"boldstart ventures","thesis":"Day one partner for developer first, crypto, and SaaS founders. We love deeply technical founders solving hard infrastructure problems.","check_size":"$1M - $3M","stage_focus":["Pre-seed","Seed"],"industry_tags":["DevTools","Infrastructure","Crypto"],"geography_focus":["Global","US"],"notable_portfolio":["Snyk","Blockdaemon","Superhuman"],"website":"https://boldstart.vc"},
  {"fund_name":"Heavybit","thesis":"The leading investor in developer-first startups. We help technical founders launch, gain traction, and build enterprise-ready companies.","check_size":"$1M - $5M","stage_focus":["Seed","Series A"],"industry_tags":["DevTools","Infrastructure","Open Source"],"geography_focus":["Global","US"],"notable_portfolio":["PagerDuty","Sanity","Netlify"],"website":"https://www.heavybit.com"},
  {"fund_name":"Amplify Partners","thesis":"We invest in technical founders building the next generation of IT infrastructure, developer tools, and data platforms.","check_size":"$2M - $8M","stage_focus":["Seed","Series A"],"industry_tags":["DevTools","Infrastructure","AI","Data"],"geography_focus":["US"],"notable_portfolio":["Datadog","OCTO","dbt Labs"],"website":"https://www.amplifypartners.com"},
  {"fund_name":"OSS Capital","thesis":"We exclusively back early-stage founders building Commercial Open Source Software (COSS) companies.","check_size":"$500k - $2M","stage_focus":["Pre-seed","Seed","Series A"],"industry_tags":["Open Source","DevTools"],"geography_focus":["Global"],"notable_portfolio":["Cal.com","Appsmith","Hoppscotch"],"website":"https://oss.capital"},
  {"fund_name":"Sequoia Capital","thesis":"We help the daring build legendary companies, from idea to IPO and beyond. Sequoia is an early-stage and growth-stage investor.","check_size":"$1M - $10M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","Enterprise","Consumer","AI"],"geography_focus":["Global"],"notable_portfolio":["Apple","Google","WhatsApp"],"website":"https://www.sequoiacap.com"},
  {"fund_name":"Andreessen Horowitz (a16z)","thesis":"We invest in software eating the world. We back bold entrepreneurs building the future through technology.","check_size":"$1M - $50M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","Crypto","Enterprise","Consumer","AI"],"geography_focus":["Global","US"],"notable_portfolio":["Facebook","Coinbase","Figma"],"website":"https://a16z.com"},
  {"fund_name":"Point Nine Capital","thesis":"We are a seed-stage venture capital firm focused on B2B SaaS and B2B marketplaces globally.","check_size":"$1M - $3M","stage_focus":["Seed"],"industry_tags":["B2B SaaS","Marketplaces"],"geography_focus":["Europe","Global"],"notable_portfolio":["Zendesk","Typeform","Docplanner"],"website":"https://www.pointnine.com"},
  {"fund_name":"Cherry Ventures","thesis":"We champion founders in Europe from their earliest days. We are generalist seed investors.","check_size":"$1M - $4M","stage_focus":["Pre-seed","Seed"],"industry_tags":["Generalist","Consumer","B2B SaaS"],"geography_focus":["Europe"],"notable_portfolio":["FlixBus","Auto1 Group","Forto"],"website":"https://www.cherry.vc"},
  {"fund_name":"First Round Capital","thesis":"We are the seed-stage firm that builds the most supportive community for founders.","check_size":"$1M - $4M","stage_focus":["Pre-seed","Seed"],"industry_tags":["Generalist","B2B SaaS","Consumer"],"geography_focus":["US"],"notable_portfolio":["Uber","Notion","Roblox"],"website":"https://firstround.com"},
  {"fund_name":"Bessemer Venture Partners","thesis":"BVP helps entrepreneurs lay strong foundations to build and forge long-standing companies.","check_size":"$1M - $20M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","Enterprise","Consumer","FinTech"],"geography_focus":["Global"],"notable_portfolio":["LinkedIn","Twilio","Shopify"],"website":"https://www.bvp.com"},
  {"fund_name":"Index Ventures","thesis":"We back the best and most ambitious entrepreneurs across all stages to build category-defining businesses.","check_size":"$1M - $20M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","FinTech","Consumer","B2B SaaS"],"geography_focus":["Europe","US","Global"],"notable_portfolio":["Dropbox","Slack","Figma"],"website":"https://www.indexventures.com"},
  {"fund_name":"Lightspeed Venture Partners","thesis":"We invest globally in enterprise, consumer, and health founders who are shaping the future.","check_size":"$1M - $25M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","Enterprise","Consumer","FinTech"],"geography_focus":["Global"],"notable_portfolio":["Snap","Rippling","MuleSoft"],"website":"https://lsvp.com"},
  {"fund_name":"Accel","thesis":"We partner with exceptional founders from inception through all phases of private company growth.","check_size":"$1M - $20M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","B2B SaaS","Consumer","DevTools"],"geography_focus":["Global"],"notable_portfolio":["Facebook","Atlassian","Spotify"],"website":"https://www.accel.com"},
  {"fund_name":"Bain Capital Ventures","thesis":"From seed to growth, we back founders building legendary infrastructure, fintech, application, and commerce companies.","check_size":"$1M - $50M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","Infrastructure","FinTech","B2B SaaS"],"geography_focus":["US","Global"],"notable_portfolio":["DocuSign","SendGrid","Redis"],"website":"https://www.baincapitalventures.com"},
  {"fund_name":"Greylock Partners","thesis":"We partner with early-stage founders to build enterprise and consumer software companies that define new categories.","check_size":"$1M - $10M","stage_focus":["Seed","Series A"],"industry_tags":["Enterprise","Consumer","Cybersecurity","AI"],"geography_focus":["US"],"notable_portfolio":["Workday","Palo Alto Networks","LinkedIn"],"website":"https://greylock.com"},
  {"fund_name":"Unusual Ventures","thesis":"We provide a breakthrough level of support for early-stage founders building enterprise tech.","check_size":"$1M - $5M","stage_focus":["Pre-seed","Seed"],"industry_tags":["Enterprise","DevTools","B2B SaaS"],"geography_focus":["US"],"notable_portfolio":["Arctic Wolf","Harness","Vivun"],"website":"https://www.unusual.vc"},
  {"fund_name":"Crane Venture Partners","thesis":"We back deep tech and enterprise founders in Europe solving hard problems with data and code.","check_size":"$1M - $4M","stage_focus":["Seed"],"industry_tags":["Enterprise","DeepTech","Data","AI"],"geography_focus":["Europe"],"notable_portfolio":["Onfido","Tessian","Forto"],"website":"https://crane.vc"},
  {"fund_name":"Founder Collective","thesis":"We are a seed-stage venture capital fund, built by founders, for founders. We back weird, wonderful, and wild startups.","check_size":"$500k - $2M","stage_focus":["Seed"],"industry_tags":["Generalist","Consumer","B2B SaaS"],"geography_focus":["US","Global"],"notable_portfolio":["Uber","Airtable","BuzzFeed"],"website":"https://www.foundercollective.com"},
  {"fund_name":"Benchmark","thesis":"We are a partnership of equal partners. We back mission-driven founders at the earliest stages and walk beside them for the long haul.","check_size":"$1M - $10M","stage_focus":["Seed","Series A"],"industry_tags":["Generalist","Marketplaces","Enterprise","Consumer"],"geography_focus":["US","Global"],"notable_portfolio":["Uber","Twitter","eBay","Snapchat"],"website":"https://www.benchmark.com"},
  {"fund_name":"Accel India","thesis":"We partner with exceptional founders from inception through all phases of private company growth in the Indian ecosystem.","check_size":"$1M - $15M","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","B2B SaaS","Consumer","FinTech","E-commerce"],"geography_focus":["India"],"notable_portfolio":["Flipkart","Swiggy","Freshworks"],"website":"https://www.accel.com/india"},
  {"fund_name":"Blume Ventures","thesis":"We are a seed and pre-seed venture fund that backs startups with both funding and active mentoring.","check_size":"$500k - $3M","stage_focus":["Pre-seed","Seed"],"industry_tags":["Generalist","B2B SaaS","Consumer","DeepTech","HealthTech"],"geography_focus":["India"],"notable_portfolio":["Unacademy","Purplle","GreyOrange"],"website":"https://blume.vc"},
  {"fund_name":"Elevation Capital","thesis":"We partner with visionary founders in India across early stages to help them build category-defining businesses.","check_size":"$1M - $10M","stage_focus":["Seed","Series A"],"industry_tags":["Generalist","Consumer","FinTech","B2B SaaS","HealthTech"],"geography_focus":["India"],"notable_portfolio":["Paytm","Swiggy","Meesho"],"website":"https://elevationcapital.com"},
  {"fund_name":"Peak XV Partners","thesis":"Formerly Sequoia India & SEA, we partner with founders across early, growth, and public stages to build enduring companies.","check_size":"$1M - $20M+","stage_focus":["Seed","Series A","Growth"],"industry_tags":["Generalist","Consumer","FinTech","B2B SaaS","DevTools","AI"],"geography_focus":["India","South Asia"],"notable_portfolio":["Zomato","Pine Labs","Cred"],"website":"https://www.peakxv.com"},
  {"fund_name":"Nexus Venture Partners","thesis":"We are a US-India venture capital firm backing extraordinary founders building product-first companies.","check_size":"$1M - $10M","stage_focus":["Seed","Series A"],"industry_tags":["B2B SaaS","Enterprise","DevTools","Consumer"],"geography_focus":["India","US"],"notable_portfolio":["Postman","Hasura","Zepto"],"website":"https://nexusvp.com"}
]

STAGE_ORDER = {"Pre-seed": 0, "Seed": 1, "Series A": 2, "Growth": 3}

def score_fund(fund, ctx):
    score = 0
    fund_tags = fund.get("industry_tags", [])
    extracted_tags = ctx.get("extracted_tags", ["Generalist"])
    tag_points = 0
    matched_tags = []
    for tag in extracted_tags:
        if tag in fund_tags:
            tag_points += 5 if tag == "Generalist" else 20
            matched_tags.append(tag)
    tag_points = min(tag_points, 60)
    score += tag_points
    stage_hint = ctx.get("stage_hint")
    fund_stages = fund.get("stage_focus", [])
    if not stage_hint:
        score += 10
    elif fund_stages:
        if stage_hint in fund_stages:
            score += 20
        elif stage_hint in STAGE_ORDER:
            hint_idx = STAGE_ORDER[stage_hint]
            if any(f in STAGE_ORDER and abs(STAGE_ORDER[f] - hint_idx) == 1 for f in fund_stages):
                score += 10
    geo_hint = ctx.get("geography_hint")
    fund_geo = fund.get("geography_focus", ["Global"])
    if not geo_hint or geo_hint == "Global":
        score += 10
    elif fund_geo == ["India"] and geo_hint == "US":
        pass
    elif geo_hint in fund_geo:
        score += 20
    elif "Global" in fund_geo:
        score += 15
    if geo_hint == "US" and "India" in fund_geo and "US" not in fund_geo and "Global" not in fund_geo:
        score = max(0, score - 30)
    if fund_tags and extracted_tags and fund_tags[0] not in extracted_tags and tag_points <= 20:
        score = max(0, score - 15)
    return score, matched_tags

scored = []
for fund in VC_FUNDS:
    score, matched_tags = score_fund(fund, context)
    tier = "High" if score >= 70 else ("Medium" if score >= 40 else "Low")
    scored.append({
        "fund_name": fund["fund_name"],
        "thesis": fund["thesis"],
        "check_size": fund["check_size"],
        "stage_focus": fund["stage_focus"],
        "industry_tags": fund["industry_tags"],
        "geography_focus": fund["geography_focus"],
        "notable_portfolio": fund["notable_portfolio"],
        "website": fund["website"],
        "source": "verified (fund website)",
        "score": score,
        "confidence": tier,
        "matched_tags": matched_tags
    })

scored.sort(key=lambda x: (-x["score"], x["fund_name"]))
relevant = [m for m in scored if m["confidence"] in ("High", "Medium")]

curated_comparables = []
for m in relevant:
    for company in m.get("notable_portfolio", []):
        if company not in curated_comparables:
            curated_comparables.append(company)

output = {
    "high_medium_matches": relevant,
    "curated_comparables": curated_comparables[:6]
}
json.dump(output, open('/tmp/vc-curated-matches.json', 'w'), indent=2)
print(f'Curated matches: {len(relevant)} High/Medium confidence funds')
for m in relevant[:8]:
    print(f'  {m["confidence"]:6} ({m["score"]:3}) {m["fund_name"]}')
print(f'Seed comparables from portfolio: {curated_comparables[:6]}')
PYEOF

---

步骤 6：通过 Tavily 发现可比公司

将步骤 5b 中整理的投资组合公司作为种子可比公司载入：

python3 -c "
import json
matches = json.load(open('/tmp/vc-curated-matches.json'))
curated = matches.get('curated_comparables', [])
print(f'Curated portfolio comparables ({len(curated)}): {curated}')
need = max(0, 5 - len(curated))
print(f'Tavily will supplement with up to {need} more')
"

不要使用 AI 的训练知识来生成可比公司。 上述整理的投资组合公司已根据经过验证的基金数据筛选，属于零幻觉可比公司。Tavily 会补充特定于 L3 细分领域的公司。

python3 << 'PYEOF'
import json, os, urllib.request

analysis = json.load(open('/tmp/vc-product-analysis.json'))
l2 = analysis['industry_taxonomy']['l2']
l3 = analysis['industry_taxonomy']['l3']
tavily_key = os.environ.get('TAVILY_API_KEY', '')

queries = [
    f'"{l3}" startup raised funding venture capital seed series',
    f'"{l2}" companies venture backed funded startup'
]

all_results = []
for query in queries:
    payload = json.dumps({
        "api_key": tavily_key,
        "query": query,
        "search_depth": "advanced",
        "max_results": 8,
        "include_answer": True
    }).encode()

    req = urllib.request.Request(
        'https://api.tavily.com/search',
        data=payload,
        headers={'Content-Type': 'application/json'},
        method='POST'
    )

    try:
        with urllib.request.urlopen(req, timeout=30) as resp:
            result = json.loads(resp.read())
            all_results.append({
                'query': query,
                'answer': result.get('answer', ''),
                'results': [
                    {'title': r.get('title',''), 'url': r.get('url',''), 'content': r.get('content','')[:500]}
                    for r in result.get('results', [])
                ]
            })
            print(f'Comparable search: {len(result.get("results", []))} results for "{query[:60]}"')
    except Exception as e:
        print(f'Comparable search FAILED: {e}')
        all_results.append({'query': query, 'answer': '', 'results': [], 'error': str(e)})

json.dump(all_results, open('/tmp/vc-comparable-search.json', 'w'), indent=2)
PYEOF

打印结果以供 AI 选择：

python3 -c "
import json
results = json.load(open('/tmp/vc-comparable-search.json'))
for r in results:
    print(f'Query: {r[\"query\"]}')
    print(f'Answer: {r.get(\"answer\",\"\")[:400]}')
    for item in r.get('results', []):
        print(f'  - {item[\"title\"]} | {item[\"url\"]}')
        print(f'    {item[\"content\"][:200]}')
    print()
"

AI 指令： 将 /tmp/vc-curated-matches.json 中整理的投资组合公司与上述 Tavily 搜索结果合并。恰好选出 5 家可比公司。优先选择整理的投资组合公司（这些公司已验证，确实是匹配 VC 基金的投资组合公司）。如有需要，再补充 Tavily 发现的公司，凑足 5 家。

为每家可比公司填写：

• name：公司名称
• similarity_reason：用一句话解释其匹配原因（整理的公司：注明投资该公司的基金；Tavily 发现的公司：引用搜索摘要）
• source_url：整理的公司填写基金投资组合网站，发现的公司填写 Tavily 结果 URL
• estimated_stage：根据整理的数据或搜索摘要填写；如果未知，则填写 "not in search data"
• source_type："curated_portfolio" 或 "tavily_discovered"

使用 comparable_companies 数组更新 /tmp/vc-product-analysis.json：

python3 << 'PYEOF'
import json

analysis = json.load(open('/tmp/vc-product-analysis.json'))

analysis['comparable_companies'] = [
    # FILL 5 companies -- curated_portfolio first, then tavily_discovered
    # Each: {"name": str, "similarity_reason": str, "source_url": str, "estimated_stage": str, "source_type": str}
]

json.dump(analysis, open('/tmp/vc-product-analysis.json', 'w'), indent=2)
print('Comparables written:', ', '.join(c['name'] for c in analysis['comparable_companies']))
PYEOF

如果搜索结果中出现的可比公司少于 3 家： 扩大查询范围。运行第三次搜索："[l1] startup" funding round venture capital。如果结果仍然有限，则使用现有结果继续，并在 data_quality_flags 中标记。

---

步骤 7：轨道 A——谁投资了可比公司

针对每家可比公司运行一次 Tavily 搜索，共运行 5 次。

python3 << 'PYEOF'
import json, os, urllib.request

analysis = json.load(open('/tmp/vc-product-analysis.json'))
comparables = analysis['comparable_companies']
tavily_key = os.environ.get('TAVILY_API_KEY', '')
all_track_a = []

for comp in comparables:
    company = comp['name']
    query = f'"{company}" investors funding venture capital backed seed series'

    payload = json.dumps({
        "api_key": tavily_key,
        "query": query,
        "search_depth": "advanced",
        "max_results": 5,
        "include_answer": True
    }).encode()

    req = urllib.request.Request(
        'https://api.tavily.com/search',
        data=payload,
        headers={'Content-Type': 'application/json'},
        method='POST'
    )

    try:
        with urllib.request.urlopen(req, timeout=30) as resp:
            result = json.loads(resp.read())
            all_track_a.append({
                'comparable_company': company,
                'similarity_reason': comp['similarity_reason'],
                'query': query,
                'answer': result.get('answer', ''),
                'results': result.get('results', [])
            })
            print(f'Track A - {company}: {len(result.get("results", []))} results')
    except Exception as e:
        print(f'Track A - {company}: FAILED ({e})')
        all_track_a.append({
            'comparable_company': company,
            'similarity_reason': comp['similarity_reason'],
            'query': query,
            'answer': '',
            'results': [],
            'error': str(e)
        })

json.dump(all_track_a, open('/tmp/vc-tracka-results.json', 'w'), indent=2)
print(f'Track A complete. Comparables with results: {sum(1 for r in all_track_a if r.get("results"))}')
PYEOF

如果所有 5 次轨道 A 搜索都没有返回结果： 使用范围更广的查询重新运行步骤 6。重试时选择媒体报道较多的知名公司。如果仍然没有结果：仅继续使用轨道 B，并在 data_quality_flags 中标记。

---

步骤 8：轨道 B——对该领域有投资论点的 VC

使用步骤 5 中的 L2 和 L3 分类运行 3 次 Tavily 搜索。

python3 << 'PYEOF'
import json, os, urllib.request

analysis = json.load(open('/tmp/vc-product-analysis.json'))
l2 = analysis['industry_taxonomy']['l2']
l3 = analysis['industry_taxonomy']['l3']
stage = analysis['detected_stage']
tavily_key = os.environ.get('TAVILY_API_KEY', '')

queries = [
    {'name': 'thesis_l3', 'query': f'venture capital investment thesis "{l3}" investing 2023 OR 2024 OR 2025'},
    {'name': 'thesis_l2', 'query': f'VC fund "{l2}" investment thesis portfolio companies'},
    {'name': 'stage_space', 'query': f'{stage} investors "{l3}" startup venture capital fund'}
]

all_track_b = []

for q in queries:
    payload = json.dumps({
        "api_key": tavily_key,
        "query": q['query'],
        "search_depth": "advanced",
        "max_results": 7,
        "include_answer": True
    }).encode()

    req = urllib.request.Request(
        'https://api.tavily.com/search',
        data=payload,
        headers={'Content-Type': 'application/json'},
        method='POST'
    )

    try:
        with urllib.request.urlopen(req, timeout=30) as resp:
            result = json.loads(resp.read())
            all_track_b.append({
                'query_name': q['name'],
                'query': q['query'],
                'answer': result.get('answer', ''),
                'results': result.get('results', [])
            })
            print(f"Track B - {q['name']}: {len(result.get('results', []))} results")
    except Exception as e:
        print(f"Track B - {q['name']}: FAILED ({e})")
        all_track_b.append({'query_name': q['name'], 'query': q['query'], 'answer': '', 'results': [], 'error': str(e)})

json.dump(all_track_b, open('/tmp/vc-trackb-results.json', 'w'), indent=2)
PYEOF

如果所有 3 次轨道 B 搜索都没有返回结果： 仅使用轨道 A 的结果继续。在 data_quality_flags 中注明："No thesis-led investors found via public search."

---

第 9 步：综合——对所有 VC 排名并评分

打印研究数据：

python3 -c "
import json

analysis = json.load(open('/tmp/vc-product-analysis.json'))
track_a = json.load(open('/tmp/vc-tracka-results.json'))
track_b = json.load(open('/tmp/vc-trackb-results.json'))
curated = json.load(open('/tmp/vc-curated-matches.json'))

track_a_summary = []
for item in track_a:
    snippets = [{'title': r.get('title',''), 'url': r.get('url',''), 'content': r.get('content','')[:400]}
                for r in item.get('results', [])[:3]]
    track_a_summary.append({
        'comparable_company': item['comparable_company'],
        'similarity_reason': item['similarity_reason'],
        'answer': item.get('answer', '')[:500],
        'top_results': snippets
    })

track_b_summary = []
for item in track_b:
    snippets = [{'title': r.get('title',''), 'url': r.get('url',''), 'content': r.get('content','')[:400]}
                for r in item.get('results', [])[:4]]
    track_b_summary.append({
        'query_name': item['query_name'],
        'answer': item.get('answer', '')[:500],
        'top_results': snippets
    })

curated_summary = []
for m in curated.get('high_medium_matches', []):
    curated_summary.append({
        'fund_name': m['fund_name'],
        'confidence': m['confidence'],
        'score': m['score'],
        'matched_tags': m['matched_tags'],
        'thesis': m['thesis'],
        'check_size': m['check_size'],
        'stage_focus': m['stage_focus'],
        'notable_portfolio': m['notable_portfolio'],
        'website': m['website'],
        'source': 'verified (fund website)'
    })

print(json.dumps({
    'product': {
        'name': analysis['product_name'],
        'description': analysis['one_line_description'],
        'industry': analysis['industry_taxonomy'],
        'icp': analysis['icp'],
        'stage': analysis['detected_stage'],
        'stage_confidence': analysis['stage_confidence'],
        'geography': analysis['geography_bias']
    },
    'curated_matches': curated_summary,
    'track_a_research': track_a_summary,
    'track_b_research': track_b_summary
}, indent=2))
"

AI 指令——零幻觉规则：

输出中的每个字段都必须能追溯到上方打印的数据。规则：

1. curated_vcs： 直接使用 curated_matches 数据。这些数据已经过预先验证，无需 Tavily 证据。fund_overview 来自精选数据中的 thesis 字段。check_size 和 stage_focus 来自精选数据中的相应字段。即使是这些基金，也不得根据训练知识填充信息。
2. VC 名称（Track A / B）： 仅在名称逐字出现在摘要文本或标题中时才纳入基金。无一例外。
3. evidence_company（Track A）： 该基金投资过的可比公司，必须在摘要文本中明确提及，不得推断。
4. thesis_source_title（Track B）： 搜索结果中显示的文章或帖子的准确标题。
5. fund_overview（Track A / B）： 仅从摘要文本中提取，最多 2 句话。如果摘要未介绍该基金，填写“搜索数据中未找到”。
6. thesis_summary： 紧贴摘要文本改写，不得添加训练知识中的背景信息。
7. check_size（Track A / B）： 仅依据摘要数据。如果未提及，填写“搜索数据中未提及”。
8. portfolio_in_space： 仅列出搜索摘要中出现的公司。如果没有，填写“搜索数据中未找到”。
9. stage_fit_score 1-10： 如果 VC 声明的阶段与检测到的产品阶段不匹配，扣 3 分。
10. space_fit_score 1-10： 仅当摘要或精选数据表明 VC 投资过 L3 细分领域的 2 家或以上公司时，才可给 9-10 分。
11. approach_method： 选用以下一种——cold email / warm intro required / AngelList / application form / Twitter/X DM。根据摘要或基金网站推断。
12. outreach_hook： 必须提及具体的投资组合信号或投资论点引述。诸如“突出你的增长表现”之类的通用切入点不可接受。
13. 不得使用 em dash。不得使用营销语言。

写入 /tmp/vc-final-list.json：

• product_summary：name、one_line_description、industry_l1、industry_l2、industry_l3、detected_stage、comparable_companies_used（仅名称）
• curated_vcs：fund_name、confidence（"High"/"Medium"）、matched_tags、fund_overview（来自 thesis 字段）、check_size、stage_focus、website、source（"verified (fund website)"）、stage_fit_score、space_fit_score
• track_a_vcs：fund_name、evidence_company（必填）、evidence_source_url、stage_focus、check_size、fund_overview、thesis_summary、stage_fit_score、space_fit_score、approach_method
• track_b_vcs：fund_name、thesis_source_title（必填）、thesis_source_url、stage_focus、check_size、fund_overview、thesis_summary、stage_fit_score、space_fit_score、approach_method
• top_5_deep_dives：fund_name、track（"Curated"/"A"/"B"）、fund_overview、why_fit、portfolio_in_space、how_to_approach（至少 30 个字符）、outreach_hook
• outreach_hooks：3 个对象，包含 hook_type、hook_text（2-3 句话）、best_for
• data_quality_flags：gaps、missing fields、low-confidence areas

python3 << 'PYEOF'
import json

result = {
    # FILL from synthesis above
    # Must include: product_summary, curated_vcs, track_a_vcs, track_b_vcs, top_5_deep_dives, outreach_hooks, data_quality_flags
}

json.dump(result, open('/tmp/vc-final-list.json', 'w'), indent=2)
print(f'Synthesis written. Curated: {len(result.get("curated_vcs", []))} VCs. Track A: {len(result.get("track_a_vcs", []))} VCs. Track B: {len(result.get("track_b_vcs", []))} VCs.')
PYEOF

---

第 10 步：自检

python3 << 'PYEOF'
import json

result = json.load(open('/tmp/vc-final-list.json'))
failures = []

# Remove Track A VCs missing evidence_company
original_a = len(result.get('track_a_vcs', []))
result['track_a_vcs'] = [v for v in result.get('track_a_vcs', []) if v.get('evidence_company')]
removed_a = original_a - len(result['track_a_vcs'])
if removed_a > 0:
    failures.append(f'Removed {removed_a} Track A VC(s) missing evidence_company')

# Remove Track B VCs missing thesis_source_title
original_b = len(result.get('track_b_vcs', []))
result['track_b_vcs'] = [v for v in result.get('track_b_vcs', []) if v.get('thesis_source_title')]
removed_b = original_b - len(result['track_b_vcs'])
if removed_b > 0:
    failures.append(f'Removed {removed_b} Track B VC(s) missing thesis_source_title')

# Remove deep dives for VCs that were stripped from all tracks
valid_funds = (
    {v['fund_name'] for v in result.get('curated_vcs', [])} |
    {v['fund_name'] for v in result.get('track_a_vcs', [])} |
    {v['fund_name'] for v in result.get('track_b_vcs', [])}
)
original_dives = len(result.get('top_5_deep_dives', []))
result['top_5_deep_dives'] = [d for d in result.get('top_5_deep_dives', []) if d.get('fund_name') in valid_funds]
removed_dives = original_dives - len(result['top_5_deep_dives'])
if removed_dives > 0:
    failures.append(f'Removed {removed_dives} deep dive(s) for funds stripped during QA')

# Check top 5 deep dives
dives = result.get('top_5_deep_dives', [])
if len(dives) < 5:
    failures.append(f'Only {len(dives)} deep dives (expected 5) -- insufficient search data')
for dd in dives:
    if not dd.get('how_to_approach') or len(dd.get('how_to_approach', '')) < 30:
        dd['how_to_approach'] = 'Approach method not determinable from search data. Check the fund website directly for application instructions.'
        failures.append(f"Fixed: '{dd.get('fund_name')}' had missing how_to_approach")
    if not dd.get('fund_overview') or dd.get('fund_overview') == '':
        dd['fund_overview'] = 'not found in search data'

# Check outreach hooks count
if len(result.get('outreach_hooks', [])) != 3:
    failures.append(f"Expected 3 outreach hooks, got {len(result.get('outreach_hooks', []))}")

# Check for em dashes
full_text = json.dumps(result)
if '—' in full_text:
    result = json.loads(full_text.replace('—', '-'))
    failures.append('Fixed: em dash characters replaced with hyphens')

# Check for forbidden words
forbidden = ['powerful', 'robust', 'seamless', 'innovative', 'game-changing', 'streamline', 'leverage', 'transform']
full_text_lower = json.dumps(result).lower()
for word in forbidden:
    if word in full_text_lower:
        failures.append(f"Warning: forbidden word '{word}' found in output -- review before presenting")

# Flag any "not found in search data" entries so user knows coverage is incomplete
not_found_count = json.dumps(result).count('not found in search data')
if not_found_count > 0:
    failures.append(f'INFO: {not_found_count} field(s) marked "not found in search data" -- verify directly before outreach')

if 'data_quality_flags' not in result:
    result['data_quality_flags'] = []
result['data_quality_flags'].extend(failures)

json.dump(result, open('/tmp/vc-final-list.json', 'w'), indent=2)
print(f'QA complete. Issues addressed: {len(failures)}')
for f in failures:
    print(f'  - {f}')
if not failures:
    print('All QA checks passed.')
PYEOF

---

第 11 步：保存并呈现输出

DATE=$(date +%Y-%m-%d)
OUTPUT_FILE="docs/vc-intel/${PRODUCT_SLUG}-${DATE}.md"
mkdir -p docs/vc-intel

呈现最终输出：

## VC Finder: [product_name]
Date: [today] | Stage: [detected_stage] ([stage_confidence] confidence) | Geography: [geography_bias]

---

### Product Analysis

What it does: [one_line_description]
Industry: [l1] > [l2] > [l3]
Buyer: [buyer_persona] at [company_type], [company_size]
Comparable companies used: [comma-separated list, noting source_type for each]

---

### Curated Matches (Verified)

*Funds matched from a verified dataset of 25 VC funds sourced from fund websites. Zero hallucination -- details come directly from the dataset.*

| Fund | Confidence | Stage Focus | Check Size | Matched Tags |
|---|---|---|---|---|
[one row per curated VC, sorted by confidence then score]

---

### Track A: VCs Who Backed Similar Companies

*These investors have already written a check in this space. Evidence from live Tavily search.*

| Fund | Backed Comparable | Stage Focus | Check Size | Fit Score | Approach |
|---|---|---|---|---|---|
[one row per Track A VC, sorted by space_fit_score descending]

---

### Track B: Thesis-Led Investors

*These investors are actively publishing about this space.*

| Fund | Thesis Source | Stage Focus | Check Size | Fit Score | Approach |
|---|---|---|---|---|---|
[one row per Track B VC, sorted by space_fit_score descending]

---

### Top 5 Deep Dives

#### [N]. [Fund Name] (Track [Curated/A/B])

Overview: [fund_overview -- from dataset or search data only]
Why it fits: [why_fit]
Portfolio in this space: [from dataset or search data, or "not found in search data"]
How to approach: [how_to_approach]
Outreach hook: "[outreach_hook]"

[repeat for all available deep dives]

---

### 3 Outreach Hooks for This Product Type

**1. [hook_type]**
[hook_text]
Best for: [best_for]

[repeat for all 3]

---
Data quality notes: [data_quality_flags, or "None"]
Saved to: docs/vc-intel/[PRODUCT_SLUG]-[DATE].md

清理临时文件：

rm -f /tmp/vc-product-raw.md /tmp/vc-stage-signals.json /tmp/vc-product-analysis.json \
      /tmp/vc-product-context.json /tmp/vc-curated-matches.json /tmp/vc-comparable-search.json \
      /tmp/vc-tracka-results.json /tmp/vc-trackb-results.json /tmp/vc-final-list.json