1 files changed, 56 insertions, 0 deletions
diff --git a/v2/README.md b/v2/README.md
new file mode 100644
index 0000000..a5b848a
--- /dev/null
+++ b/v2/README.md
@@ -0,0 +1,56 @@
+# License Classifier v2
+
+This is a substantial revision of the license classifier with a focus on improved accuracy and performance.
+
+## Glossary
+
+- corpus dictionary - contains all the unique tokens stored in the corpus of
+documents to match. Any tokens in the target document that aren't in the corpus
+dictionary are mapped to an invalid value.
+
+- document - an internal-only data type that contains sequenced token information
+for a source or target content for matching.
+
+- source content - a body of text that can be matched by the scanner.
+
+- target content - the argument to Match that is scanned for matches with source
+content.
+
+- indexed document - an internal-only data type that maps a document to the
+corpus dictionary, resulting in a compressed representation suitable for fast
+text searching and mapping operations. an indexed document is necessarily
+tightly coupled to its corpus.
+
+- frequency table - a lookup table holding per-token counts of the number of
+times a token appears in content. used for fast filtering of target content
+against different source contents.
+
+- q-gram - a substring of content of length q tokens used to efficiently match
+ranges of text. For background on the q-gram algorithms used, please see
+[Indexing Methods for Approximate String Matching](https://users.dcc.uchile.cl/~gnavarro/ps/deb01.pdf)
+
+- searchset - a data structure that uses q-grams to identify ranges of text in
+the target that correspond to a range of text in the source. The searchset
+algorithms compensate for the allowable error in matching text exactly, dealing
+with additional or missing tokens.
+
+
+## Migrating from v1
+
+The API for the classifier versions is quite similar, but there are two key
+distinctions to be aware of while migrating usages.
+
+The confidence value for the v2 classifier is applied uniformly to results; it
+will never return a match that is lower confidence than the threshold. In v1,
+MultipleMatch behaved this way, but NearestMatch would return a value
+regardless of the confidence match. Users often verified that the confidence
+was above the threshold, but this is no longer necessary.
+
+The second change is that the classifier now returns all matches against the
+supplied corpus. The v1 classifier allowed filtering on header matches via a
+boolean field. This can be emulated by creating a license classifier with a
+reduced corpus if matching against headers is not desired. Alternatively, the
+user can use the MatchType field in the Match struct to filter out unwanted
+matches.
+
+