How I Built Smart Search Across 3,500 Contacts in 2 Hours

Three thousand five hundred names in a file. Before: boolean logic, jq, trial and error. After: a sentence.

  classic search:            semantic search:
  ──────────────             ───────────────
  city == "Berlin"           "founders in Berlin
  AND role == "founder"       working on AI"
  AND tags includes "AI"
        ↓                           ↓
  what you wrote              what you meant

The filter knows only what you typed. The vector knows what you meant.

---

The word “product” does not appear in the result. The meaning of “product” does.

  query: "people who build products"

  found:
  ──────
  "founder of startup X"
  "building my own project"
  "indie hacker"
  "launching my own service"

Letters point to other letters. Embeddings point to meaning.

---

  what it cost to build:

  setup time        ──  2 hours
  indexing time     ──  90 seconds
  embedding cost    ──  < $0.01
  Qdrant Cloud      ──  free tier

The infrastructure is not the hard part. The hard part was the raw data — three years of contacts in a flat file, no structure, no roles, no cities.

Enrichment turned noise into signal. Indexing turned signal into searchable.

---

  Mac
  ┌──────────────┐      ┌──────────────────┐
  │Claude Desktop│ ───▶ │  MCP server      │
  └──────────────┘      └────────┬─────────┘
                                 │
                    ┌────────────┴────────────┐
                    ▼                         ▼
          ┌──────────────────┐   ┌────────────────────┐
          │  Qdrant Cloud    │   │  Embedding API     │
          └──────────────────┘   └────────────────────┘

The server runs where you are. The data lives where it is safe. The query travels between them in milliseconds.

---

The CRM was a file I queried. Now it is a conversation I have.

The database did not change. The interface to it did.

That is the only change that matters to the person asking the question.

— Ilao Dzindin