M|BOX

The M|BOX Python SDK can be used to build deterministic customer identity resolution tools for OpenAI Agents.

M|BOX indexed 10,000 customer records in under 0.2 seconds, which the tutorial calls "sub-millisecond indexing".

M|BOX's index schema classifies single-word attributes (such as first name, last name, city) as term fields, multi-word names (such as company) as phrase fields, and customer ID and email as non-searchable fields.

The resolveCustomer function takes four parameters—first name, last name, company, city—internally calls index.match to return the top three candidates with their field-level scores, and empty-string fields are excluded from the total score calculation.

In a test that used a typo (Catherine Vos at Kestrel Manufacturing) as the search, M|BOX matched Kestrel within a few milliseconds with a total score of 76, of which first name scored 100, company 90, last name 39, and city 0 with no penalty.

In the spelling-difference scenario, asking about Elizabeth Turner at Orion Systems (S spelling) in Frankfurt, the Agent successfully matched the record stored in the database with a Z spelling, customer ID c17505, with a score of 92, the M|BOX search took about 3 milliseconds, and the entire Agent run took several seconds.

In the no-such-person scenario, asking about Laura Chen at Brightwave Corporation, the tool returned zero candidates, and the Agent clearly reported that Laura Chen is not in the CRM, with no hallucination.

Vector nearest-neighbor search will always return a match, whereas M|BOX never forces a match when the data is not close enough, and this deterministic honesty carries straight through to the Agent.

The architectural value lies in the division of responsibilities: M|BOX handles deterministic matching, the model handles conversational guidance, and field-level scoring provides explainability.