Lead Enrich

Lead Enrich

Research a person (name + company) using Serper.dev Google Search. Uses the agent's built-in model for extraction. No external LLM API needed.

Works standalone — just a Serper key. Pairs with Outreach Magic for credit-saving dedup (skip Serper when leads already have LinkedIn/email) and persistent storage across sessions.

Credit-saving (with OM): checks outreachmagic first. If the lead already has LinkedIn + email at the same company, zero Serper credits spent. LinkedIn without email skips Serper — use email-finder for trykitt. Email-only records still get LinkedIn searches (1–2 Serper credits). Name matches at a different company return ambiguous so APIs aren't wasted.

Setup

Standalone (no OM)

Just a Serper key. Run serper-search directly. Results print as JSON.

python3 scripts/enrich.py serper-search --query '"Acme Corp" official website'
python3 scripts/enrich.py serper-search --query 'site:linkedin.com/in Jane Doe Acme Corp'
# → pipe results to your model for extraction, or use serper-format

Set SERPER_API_KEY in your environment or via ~/.hermes/skills/lead-enrich/config.json.

With Outreach Magic (dedup + save)

Adds pre-flight dedup (skip leads already in OM) and saves structured enrichment to your local SQLite pipeline. Requires outreachmagic skill with pipeline.py login.

python3 scripts/enrich.py check "Jane Doe" "Acme Corp"        # 0 credits
python3 scripts/enrich.py batch-check --workspace W input.json # batch dedup

Save your Serper key in Outreach Magic portal → Settings → API Keys, then run pipeline.py sync-secrets. Verify with enrich.py config (serper_api_key_source should be agent_secrets).

New install? Run the doctor:

python3 scripts/enrich.py doctor

Checks the full chain: key presence, key validity (test query), OM install, credential files, and portal sync freshness. Tells you exactly what's missing.

3. Email finding (email-finder skill)

Email find is not in lead-enrich v2+. Install email-finder with outreachmagic using the platform install guide. Or install standalone via npx skills add outreachmagic/email-finder.

After Serper enrichment saves company_domain, run:

python3 ~/.hermes/skills/email-finder/scripts/email_finder.py find --name "Jane Doe" \
  --domain acme.com --linkedin "https://linkedin.com/in/janedoe" --save

See references/email-finder.md and the email-finder skill docs.

Workspace rollups (no Serper credits): after saving leads, use outreachmagic workspace summary --workspace <slug> --json for tag counts and LinkedIn connection accepted per sender. On large workspaces (>2,000 leads), add --tags-only for faster tag counts. Local DB only — pull optional.

Common workflows

Sales Nav / Vayne CSVs: use outreachmagic import-profiles --file … (auto-detects columns).

CSV / award-list workflow (preferred for 10+ people)

Paths like outreachmagic/imports/awards.csv are relative to your workspace directory (where the agent runs the command), not the skill install folder.

# 0 credits — dedup entire file first (auto-stamps serper_attempted on complete rows)
python3 scripts/enrich.py batch-check --workspace your_workspace outreachmagic/imports/awards.csv

# Re-run dedup skipping leads already tagged serper_attempted
python3 scripts/enrich.py batch-check --workspace your_workspace --skip-tagged outreachmagic/imports/awards.csv

# Serper only for rows that need LinkedIn/domain (skip team_award, exists_linkedin_*, skipped_serper_attempted)

# After research — patch title/industry only (0 Serper credits)
python3 scripts/enrich.py backfill --fields title,industry --workspace your_workspace outreachmagic/imports/patch.csv

batch-check accepts .json or .csv. backfill requires email or linkedin per row; uses chunked import-profiles via companion_common (200 rows/chunk, up to 300s/chunk; fills empty fields; add --overwrite to replace).

When to Use

• User says "research this person" / "look up Jane Doe at Acme"
• User wants to enrich a list of prospects before outreach
• User asks "do we already have this lead?" before researching
• User provides a CSV/JSON of names and companies to enrich in bulk
• User mentions Serper, lead enrichment, or person research

Agent Behavior Rules (Important)

1. Dedup first (with OM). Before any Serper API call, run enrich.py check. If the lead has LinkedIn + email at the same company, skip Serper entirely. If LinkedIn exists but no email, skip Serper and offer email-finder when the user wants an address. Never spend Serper credits on leads already complete in outreachmagic. Standalone: skip check — run serper-search directly.
2. Serper only. Prefer enrich.py serper-search --query "..." (stdlib HTTP, key from config/env). Or use curl with $SERPER_API_KEY — never embed the key in chat logs. Never scrape Google or LinkedIn directly.
3. Built-in model only. You (the agent) extract JSON from Serper results. No external LLM APIs (no Gemini, no OpenAI) — your own reasoning is the extraction engine.
4. Complete research before saving. Run the full search ladder first, then save once.
5. Save via outreachmagic (with OM). Use import-profiles for leads with LinkedIn. Always append serper_attempted to tags on save (included automatically in map-to-om output). For leads without LinkedIn but with a known lead_id, stamp the tag via stamp-attempted or import-profiles with id + tags — do not rely on notes alone. For read-only dedup checks use pipeline.py query or enrich.py check — never raw INSERT/UPDATE. Never run both save paths for the same person. Standalone: save JSON output to a file — no OM needed.
6. Tag after enrichment (with OM). Every lead that goes through Serper must get serper_attempted on save. Prevents re-processing on future runs.
7. Check tag before Serper (with OM). Before spending Serper credits, check for serper_attempted (via enrich.py check --skip-tagged or skip_reason in check output). If present and LinkedIn is still empty, skip unless the user explicitly wants a retry (e.g. stale >30 days).
8. Transparency. Show which Serper queries ran, confidence, and what was saved. The user should see exactly where their credits went.
9. Batch wisely. Cap at 50 people per run. For CSV/award lists run batch-check once on the whole file (JSON or CSV) before any Serper. Process Serper only for statuses that need LinkedIn/domain. Skip team_award, exists_linkedin_email, and skipped_serper_attempted rows.
10. Email finding — use email-finder. After enrichment saves company_domain, hand off to email-finder (email_finder.py find --save). See references/email-finder.md. Never fabricate or pattern-guess emails in this skill.

Quick Start

# Diagnose your setup (keys, OM, credential files)
python3 scripts/enrich.py doctor

# Single person (most common)
python3 scripts/enrich.py check "Jane Doe" "Acme Corp"
# → if "not_found", proceed with Serper search pack below

# With workspace (associates lead with your pipeline workspace)
python3 scripts/enrich.py check --workspace your_workspace "Jane Doe" "Acme Corp"

# Batch from JSON or CSV (run this before Serper on lists)
python3 scripts/enrich.py batch-check --workspace your_workspace outreachmagic/imports/people.json
python3 scripts/enrich.py batch-check --workspace your_workspace --skip-tagged outreachmagic/imports/awards.csv

# Stamp serper_attempted after failed LinkedIn lookup (when lead_id is known)
python3 scripts/enrich.py stamp-attempted --workspace your_workspace --lead-ids 42,43 \
  --notes "No LinkedIn found via Serper"

# Backfill title/industry on existing leads (linkedin or email required)
python3 scripts/enrich.py backfill --fields title,industry outreachmagic/imports/patch.csv

# Update skill safely from GitHub release (checksum-verified)
python3 scripts/enrich.py update --check
python3 scripts/enrich.py update

update verifies SHA256 checksums from update-manifest.json before replacing files. If checksums are missing or mismatched, the update aborts.

Core Workflow

Phase 1 — Dedup Check (0 credits)

For each person, run:

# Without workspace (org-wide lookup)
python3 scripts/enrich.py check "Jane Doe" "Acme Corp"

# With workspace (scoped to your pipeline workspace)
python3 scripts/enrich.py check --workspace your_workspace "Jane Doe" "Acme Corp"

Output per person:

Check output includes tags and optional skip_reason (has_linkedin or skipped_serper_attempted). Uses batch-lead-lookup (local, zero Serper credits).

Phase 2 — Serper Search Pack

Only run for people who need it. 2–4 searches per person depending on result quality:

Serper Credit Budget Estimator

• Per person minimum: 0 Serper credits (exists_linkedin_email).
• Common path: 2 credits (2a strict company + 2c LinkedIn profile).
• Per person hard max: 4 credits when all fallbacks are needed (2a + 2b + 2c + 2e).
• Batch formula: min=0, max=4*N where N is people in the run.
• Batch cap example: with N=50, hard max is 200 credits (worst case).

2a. Company discovery — strict (always)

Preferred (no key in shell history):

python3 scripts/enrich.py serper-search --query '"Acme Corp" official website' --label company_discovery_strict

Or retry with a simpler query:

python3 scripts/enrich.py serper-search --query 'Acme Corp website' --label company_discovery_broad

2b. Company discovery — broad (conditional)

Run only if 2a returns no organic results with an http:// or https:// link:

python3 scripts/enrich.py serper-search --query 'Acme Corp official website' --label company_discovery_broad

(Same template, unquoted company name.)

2c. LinkedIn profile (always)

Build query (unquoted company — matches variant employer names in snippets):

site:linkedin.com/in {First Last} {up to 5 words of role} {Company Name}

Example:

python3 scripts/enrich.py serper-search \
  --query 'site:linkedin.com/in Jane Doe VP Marketing Acme Corp' \
  --label linkedin_profile

2d. Retry with a different Serper query (when results are empty)

If a strict query returns no useful results, try a simpler/broader one:

python3 scripts/enrich.py serper-search --query 'Acme Corp about us' --label company_discovery_about

Stay within Serper. Never use a browser, never ask the user to manually provide a domain, and never fabricate one. If Serper returns nothing, report confidence: low with an empty company_domain.

Phase 3 — Model Extraction

Pass the formatted Serper results to yourself (the agent model) with this system instruction:

You are a research assistant. The user message contains Serper.dev Google search
results in labeled sections. You do NOT have live web search — use only the pasted
blocks.

Task: for ONE specific person at ONE company, extract:
- company_domain: registrable hostname only (no path, no www), or empty string
- company_website: full https:// homepage URL if supported, or empty string
- linkedin_url: https://linkedin.com/in/… for the named person at this company, or empty string
- confidence: high | medium | low
- note: optional explanation of ambiguity or gaps

Rules:
- Every non-empty URL must appear in the Serper blocks (minor query-string normalization ok)
- Name on LinkedIn may differ from input (nickname vs legal name) — match on same human + employer
- Before accepting `linkedin_url`, verify match quality:
  - Extract first + last name tokens from `Full name: {full_name}` (ignore punctuation)
  - The chosen `/in/` result must have *both* tokens present in either the result `title` or `snippet`
  - If no `/in/` result meets the token requirement, return `linkedin_url` as an empty string and set `confidence` to `low`
- Company public site may use a different banner name than input — prefer official evidence
- When resolving `company_domain`:
  - Prefer `knowledgeGraph.website` when present
  - Reject known aggregator/registry/catalog sites, even when returned by Serper's
    knowledge graph. These are never the company's own domain:
    - Business directories: `naceweb.org`, `usnews.com`, `niche.com`, `zoominfo.com`,
      `bloomberg.com`, `craft.co`, `owler.com`, `crunchbase.com`
    - Review sites: `yelp.com`, `g2.com`, `trustpilot.com`, `capterra.com`
    - Social media: `linkedin.com`, `facebook.com`, `instagram.com`, `twitter.com`
    - Reference: `wikipedia.org`, `britannica.com`
    - Job boards: `indeed.com`, `monster.com`, `glassdoor.com`
    - News aggregators: `prnewswire.com`, `businesswire.com`, `marketwatch.com`
    - Generic email: `gmail.com`, `yahoo.com`, `hotmail.com`
  - If `company_name` appears to be missing or matches the person name closely, return empty string
- Never fabricate URLs or slugs

Respond ONLY with a single JSON object (no markdown fences):
{"company_domain":"","company_website":"","linkedin_url":"","confidence":"medium","note":""}

Use this user message template:

### Target person
Full name: {full_name}
Stated role/title: {stated_role}
Company (as provided): {company_name}

### Search results
{formatted_serper_sections}

### Task
Return the JSON object described in the system instruction.

Then parse your own response: strip markdown fences, extract the JSON object.

LinkedIn harvest fallback: if linkedin_url is empty after extraction, scan the raw Serper organic results for /in/ URLs where the title contains both first and last name tokens. Prefer matches whose snippet/title also mention the employer/company.

Phase 4 — Save via outreachmagic

Map extracted fields to outreachmagic:

If LinkedIn found:

Every import from this skill sets --source lead_enrich and --source-detail "lead-enrich" by default. If an import_name is provided, detail appends as "lead-enrich/{import_name}".

# Org-wide (no workspace)
python3 {outreachmagic_home}/scripts/pipeline.py import-profiles \
  --source lead_enrich --source-detail "lead-enrich" \
  --json '[{"name":"Jane Doe","company":"Acme Corp","job_title":"VP Marketing","linkedin":"linkedin.com/in/janedoe","company_domain":"acme.com","tags":["nace","serper_attempted"]}]'

# Scoped to a workspace
python3 {outreachmagic_home}/scripts/pipeline.py import-profiles \
  --workspace your_workspace \
  --source lead_enrich --source-detail "lead-enrich" \
  --json '[{"name":"Jane Doe","company":"Acme Corp","job_title":"VP Marketing","linkedin":"linkedin.com/in/janedoe","company_domain":"acme.com","tags":["nace","serper_attempted"]}]'

If no LinkedIn, no email: When lead_id is known (from batch-check), stamp attempt state — do not bury failure only in notes:

python3 scripts/enrich.py stamp-attempted --workspace your_workspace --lead-ids 42 \
  --notes "No LinkedIn found via Serper"

Or via import-profiles when you also have name + company:

python3 {outreachmagic_home}/scripts/pipeline.py import-profiles \
  --workspace your_workspace \
  --source lead_enrich --source-detail "lead-enrich/no-linkedin" \
  --json '[{"id":42,"name":"Jane Doe","company":"Acme Corp","tags":["nace","serper_attempted"],"notes":"No LinkedIn found"}]'

Without a lead_id, use add-lead with notes (last resort) or report unsaved.

Email finding (email-finder skill)

After Phase 4 save, if the user wants an email and company_domain is known, use the email-finder companion — not this skill:

python3 ~/.hermes/skills/email-finder/scripts/email_finder.py find \
  --name "Jane Doe" --domain acme.com \
  --linkedin "https://linkedin.com/in/janedoe" --save

See references/email-finder.md and email-finder's email-finding-research.md.

Phase 5 — Report

Summarize per person:

Jane Doe @ Acme Corp
  ✅ Company: acme.com | https://acme.com
  ✅ LinkedIn: linkedin.com/in/janedoe
  🟢 Confidence: high
  💾 Saved to outreachmagic (lead #42)
  🔍 Serper: 2 queries
  📧 Email: use email-finder if needed

Email-only mode

When contacts are already enriched (have company_domain + linkedin in CSV or outreachmagic) but lack email:

1. Run batch-check — process only exists_linkedin_no_email (and optionally not_found rows that already have domain in the file).
2. Skip Serper entirely (0 Serper credits).
3. Run email-finder batch-find for each person with domain + LinkedIn.
4. Respect trykitt_attempted tag — email-finder skips rows already tagged.
5. Respect serper_attempted tag — lead-enrich skips rows already researched.

Useful after a prior enrichment pass or when importing a pre-researched list.

Input Formats

Single person

Research Jane Doe, VP Marketing at Acme Corp

Or with workspace:

Research Jane Doe at Acme Corp --workspace your_workspace

Batch (JSON file or inline)

{
  "people": [
    {"full_name": "Jane Doe", "company_name": "Acme Corp", "stated_role": "CEO"},
    {"full_name": "John Smith", "company_name": "Beta Inc"}
  ],
  "workspace": "your_workspace",
  "tags": ["nace"],
  "import_name": "NACE 2026 attendee"
}

Max 50 people per run.

Config Reference

config.json (copy from config.example.json):

Enrichment query patterns (SQL)

Use via pipeline.py query --sql '…' --params '[…]' --json.

Leads that still need Serper (no LinkedIn, not yet attempted):

SELECT l.id, l.name, l.company
FROM leads l
JOIN workspace_lead_tags n ON n.lead_id = l.id AND n.tag = ?
JOIN workspaces w ON n.workspace_id = w.id
WHERE w.slug = ?
  AND (l.linkedin_url IS NULL OR l.linkedin_url = '')
  AND l.id NOT IN (
    SELECT lead_id FROM workspace_lead_tags
    WHERE tag = 'serper_attempted' AND workspace_id = w.id
  )

Params: ["nace", "your_workspace"]

Enrichment attempted but failed (retry-eligible when stale):

SELECT l.id, l.name, l.company, l.updated_at
FROM leads l
JOIN workspace_lead_tags s ON s.lead_id = l.id AND s.tag = 'serper_attempted'
JOIN workspaces w ON s.workspace_id = w.id
WHERE w.slug = ?
  AND (l.linkedin_url IS NULL OR l.linkedin_url = '')
  AND l.updated_at < datetime('now', '-30 days')

Params: ["your_workspace"]

What this skill does NOT do

• ❌ Call Outreach Magic person-research API (/v1/person-research)
• ❌ Use external LLM APIs (Gemini, OpenAI, etc.) for extraction
• ❌ Scrape HTML pages or LinkedIn directly
• ❌ Guess email addresses (use email-finder skill)
• ❌ Write raw SQL or ad-hoc DB mutations (use import-profiles / add-lead; reads via pipeline.py query)
• ❌ Upload to remote servers (local-only by default)

Platform Support

Hermes: Real files live under ~/.hermes/skills/. Each profile uses symlinks only (profiles/<name>/skills/lead-enrich → ../../../skills/lead-enrich). Do not copy the skill into a profile directory.

Learn more at outreachmagic.io.

Troubleshooting