Match Company

match_company (atomic)

Resolves companies via Onfire's Matchbox2 engine. Vector search + AI matching, so partial / fuzzy / messy names work.

When to use this

• You have a company name, website, or LinkedIn URL and need the verified LinkedIn URL.
• You need to normalize or de-dupe a CRM list against a canonical identity.
• You're about to call ai_prospecting and don't have a company LinkedIn URL yet — call this first.

Skip this if the user already handed you a clean company LinkedIn URL and isn't asking you to verify it.

Input shape

match_company(companies=[
    {"names": ["Acme Inc"]},
    {"websites": ["techcorp.com"]},
    {"linkedin_urls": ["https://linkedin.com/company/acme"]},
    {"names": ["StartupCo", "Startup Co."], "websites": ["startup.io"]},  # multi-signal
])

Each entry can mix names, websites, and linkedin_urls (all are arrays — accept multiple variants per field). At least one signal is required. More signals → better match accuracy. If the user gave you both a name and a website, pass both.

Batch cap: 100 entries per call. For lists longer than 100, chunk them.

Output shape

{
  "total_count": N,
  "matched_count": M,
  "results": [
    {
      "matched": true,
      "name": "Acme Inc",
      "website": "acme.com",
      "linkedin_url": "https://linkedin.com/company/acme",
      "linkedin_id": "...",
      "size": "1001-5000",
      "found_employee_count": 1247,
      "cosine_similarity": 0.94,
      "match_type": "semantic_match",
      "match_reason": "..."
    },
    ...
  ]
}

Results are returned in the same order as the input — index alignment matters when you batch.

Hard rules

• Batch ≤ 100 per call.
• At least one of names / websites / linkedin_urls per entry.
• Don't silently propagate low-confidence matches. If matched: false or cosine_similarity looks weak (rough threshold ~0.7, but read match_reason too), surface it: "I think you mean X but I'm only ~60% sure — confirm or pick another?" Wrong identity here corrupts every downstream prospecting result.

Common pitfalls

• Sending one signal when you have two. If the user mentioned both name and website, putting them in the same entry (not two separate entries) gives a much better match.
• Treating matched: false as a hard failure. It usually means weak signal — try again with another variant of the name, or ask the user.
• Forgetting same-order alignment. When you batch 80 rows from a CSV, the result array maps 1:1 to the input — don't reshuffle without the index.

Worked examples

Single name → LinkedIn URL.

match_company(companies=[{"names": ["Pathwatch"]}])

Use the returned linkedin_url as the input to ai_prospecting.

CSV of 40 accounts (name + website columns).

match_company(companies=[
    {"names": [row["company_name"]], "websites": [row["website"]]}
    for row in csv_rows
])

After the call, filter results to matched: true for downstream calls; surface the unmatched rows back to the user so they can fix or drop them.

Verifying a LinkedIn URL.

match_company(companies=[{"linkedin_urls": ["https://linkedin.com/company/acme-inc"]}])

Use the returned firmographics to confirm the user is on the right entity (size, region) before scoring.

What this skill does NOT do

• It doesn't run prospecting — that's ai-prospecting.
• It doesn't resolve people — that's match-person.
• It doesn't enrich emails / phones — that's contact-data-enrichment.