A comprehensive phylogeny of papilionoid legumes was inferred from sequences of 2228 taxa in GenBank release 147. A semiautomated analysis pipeline was constructed to download, parse, assemble, align, combine, and build trees from a pool of 11,881 sequences. Initial steps included all-against-all BLAST similarity searches coupled with assembly, using a novel strategy for building length-homogeneous primary sequence clusters. This was followed by a combination of global and local alignment protocols to build larger secondary clusters of locally aligned sequences, thus taking into account the dramatic differences in length of the heterogeneous coding and noncoding sequence data present in GenBank. Next, clusters were checked for the presence of duplicate genes and other potentially misleading sequences and examined for combinability with other clusters on the basis of taxon overlap. Finally, two supermatrices were constructed: a "sparse" matrix based on the primary clusters alone (1794 taxa x 53,977 characters), and a somewhat more "dense" matrix based on the secondary clusters (2228 taxa x 33,168 characters). Both matrices were very sparse, with 95% of their cells containing gaps or question marks. These were subjected to extensive heuristic parsimony analyses using deterministic and stochastic heuristics, including bootstrap analyses. A "reduced consensus" bootstrap analysis was also performed to detect cryptic signal in a subtree of the data set corresponding to a "backbone" phylogeny proposed in previous studies. Overall, the dense supermatrix appeared to provide much more satisfying results, indicated by better resolution of the bootstrap tree, excellent agreement with the backbone papilionoid tree in the reduced bootstrap consensus analysis, few problematic large polytomies in the strict consensus, and less fragmentation of conventionally recognized genera. Nevertheless, at lower taxonomic levels several problems were identified and diagnosed. A large number of methodological issues in supermatrix construction at this scale are discussed, including detection of annotation errors in GenBank sequences; the shortage of effective algorithms and software for local multiple sequence alignment; the difficulty of overcoming effects of fragmentation of data into nearly disjoint blocks in sparse supermatrices; and the lack of informative tools to assess confidence limits in very large trees.