Thanks. Yes, all characters are acutally indexed, with some logic to split intelligently when required. It will always go for the most exact match first though.
I have started looking at ranking the main implementation over usage, but found the results were less useful generally. This may have just been me though, as I wanted to see usage rather then the implementation, since if I know that I will just go to the source.
Perhaps an option to request that the orginal source is given greater weight on an advanced search page would solve this.
This could become a useful code search option.
I would request that relevancy should take into account provenance. Meaning a search for pthread_t would return pthread implementations over uses.
This obviously can't make use of traditional tfidf for scoring.