You can
a) improve the efficiency of the existing structure – as other say, convert and recompile gm B into RPGLE for an immediate gain.
THen you can alter to a bound call, for another gain. Then you could look at

b) restructuring the code.

The first thing to do (even in RPG/400) is to replace CHAIN with SETLL – huge, huge, performance gain.
User OPEN FILE B and explicitly leave it open. you say you are already not setting LR on. Ensure that PGMB remains in storage.
If File A is keyed, consider if sequential access would be faster

c) restructuring the environment

If PGM A is running a primary read on FILEA, then calling PGM B to chain to FILE B, then we can infer that File B has a relationship to FILE A. Can you provide an access path joined across both, perhaps with a select on FILE B data and read THAT as a single operation?
Or run the join dynamically with SQL, or even good old Query to generate an output file of ‘only’ records where FILE B exists? – building an access path takes place down in the MI layer, and is much more efficient than RPG

You say you want PGM B to run ‘like an API’ as though you foresee its re-use elsewhere. Read about service programs and modules, and write it in a form for re-use – as a bound procedure.