How much can you truncate a SHA1 hash and be reasonably sure of having an unique ID?

Adapting the formulas on on wikipedia for the Birthday problem, you can approximate the probability of collision as 1 - e^(-n^2/(2^(b+1))), where n is the document count and b is the number of bits. Graphing this formula with n=100,000, it looks like you'll want b > 45 at least. I'd be more inclined to go with 64 to make it a nice and round number. That said, do have a plan to deal with collisions if they occur (maybe alter the timestamp slightly, or add a nonce?)

For that matter, if the sha1 is based on more than just the content of the document, why not simply make it a random ID? In this case collisions are less of a problem, as you can always generate a new random number and try again (the probability of a collision with a single try is the same, however).

There really isn't a value for this; part of what makes SHA a good general-purpose hashing algorithm is that similar data does not necessarily produce similar hashed values. Your best bet (without knowing anything else about your system) would just be to search the list of documents whose hashes start with the value supplied by the user, then either present them with a list of documents to select from or go directly to the document if there's only one.

Be careful of truncation as there is no reduction in proof that the smaller hash is secure. See Kelsey's http://csrc.nist.gov/groups/ST/hash/documents/Kelsey_Truncation.pdf. Kelsey gives to heuristic arguments stating the same ("Related Hash Outputs" and "Near Collisions"). Biham/Chen offer examples of Near Collisions; and Knudsen demonstrates Truncated Differentials.

In the end, you probably want to feed your data into an HMAC with the truncated size (the size is digested by the HMAC, too) and then use the truncated HMAC.