All hash algorithms can produce duplicates. You are taking something with thousands or potentially millions of bytes, and squashing it down to a very small number of bytes, four to be precise. Those four bytes can only contain just over 4 billion different values, so if you hash more than four billion different documents then you are guaranteed to have a duplicate. If you hash a million documents, you have a 1 in 4,000 chance of a duplicate.
The only way to reduce the likelihood of a clash is to use a longer hash. CRC32 generates a 32 bit hash, MD5 generates a 128 bits. You might think that that only increases the range by a factor of four, but in fact it is much, much, much bigger than that. You go from a 1 in 4 billion chance of a clash to a 1 in 340 million million million million million million. So long as your source value is more than 128 bits, you would need to generate a million million million hashes per second for the entire existence of the universe so far before you had a noticeable chance of a clash.
CRC64 may be enough for you, that would reduce the chance of a clash to 1 in a million million million.
Note that I said "reduce". The very nature of hashing means that a clash is possible, it's just a case of making it sufficiently unlikely.
