We use a non-Teradata data mining application (SPSS Clementine) as an front-end tool for predictive modeling and forecasting customer behaviour. We use this tool to construct our data processing queries and score our predictive models.
It automatically creates SQL that is run by Teradata. The SQL Clementine generates is not always the best or optimised, but we generaly have very good performance and obtain great results. The important thing is to construct your queries sensibly and ensure tables are indexed etc. In most cases nearly all the processing is performed by Teradata itself, and this affords us very scaleable analysis. We often forecast for our entire customer base (in the millions).
On occaision we do need to create a temporary sample file for predictive model building (creating our forercast models). A wholly Teradata solution probably wouldn't need to extract text file samples, but we find the interface easy to use and offers flexiblity if we need to work with other data formats from external sources (Excel, Access, Text etc).
Theres Teradata Warehouse Miner, which is a full data mining application developed by Teradata that operates entirely within Teradata by generating optimised SQL which is submitted. TWM allows users to do all the data exploration, preparation, modelling and scoring within the database. The modelling techniques available are Clustering, Linear Regression, Logistic Regression, Decision Trees, Neural Networks and Association/Sequence (Market basket style analysis). You can find out more about this here http://www.teradata.com/t/page/44097/index.html