diff --git a/Project.toml b/Project.toml
index d12521b..01f4198 100644
--- a/Project.toml
+++ b/Project.toml
@@ -14,7 +14,7 @@ ThreadsX = "ac1d9e8a-700a-412c-b207-f0111f4b6c0d"
 [compat]
 AbstractAlgebra = "0.15, 0.16"
 GroupsCore = "^0.3"
-KnuthBendix = "^0.2.0"
+KnuthBendix = "^0.2.1"
 OrderedCollections = "1"
 ThreadsX = "^0.1.0"
 julia = "1.3, 1.4, 1.5, 1.6"
diff --git a/README.md b/README.md
index fe6af31..7a64708 100644
--- a/README.md
+++ b/README.md
@@ -2,6 +2,152 @@
 [![Build Status](https://travis-ci.org/kalmarek/Groups.jl.svg?branch=master)](https://travis-ci.org/kalmarek/Groups.jl)
 [![codecov](https://codecov.io/gh/kalmarek/Groups.jl/branch/master/graph/badge.svg)](https://codecov.io/gh/kalmarek/Groups.jl)
 
-A very rudimentary implementation of finitely-presented groups (syllable representation). Relatively complete are only [automorphism groups of free groups](https://github.com/kalmarek/Groups.jl/blob/master/src/AutGroup.jl) and [wreath products](https://github.com/kalmarek/Groups.jl/blob/master/src/WreathProducts.jl) (which are not finitely-presented, but based on the standard normal form).
+An implementation of finitely-presented groups together with normalization (using Knuth-Bendix procedure).
 
-Have a look into `test` directory for eample use.
+The package implements `AbstractFPGroup` with three concrete types: `FreeGroup`, `FPGroup` and `AutomorphismGroup`. Here's an example usage:
+
+```julia
+julia> using Groups, GroupsCore
+
+julia> A = Alphabet([:a, :A, :b, :B, :c, :C], [2, 1, 4, 3, 6, 5])
+Alphabet of Symbol:
+        1.      :a = (:A)⁻¹
+        2.      :A = (:a)⁻¹
+        3.      :b = (:B)⁻¹
+        4.      :B = (:b)⁻¹
+        5.      :c = (:C)⁻¹
+        6.      :C = (:c)⁻¹
+
+julia> F = FreeGroup(A)
+free group on 3 generators
+
+julia> a,b,c = gens(F)
+3-element Vector{FPGroupElement{FreeGroup{Symbol}, KnuthBendix.Word{UInt8}}}:
+ a
+ b
+ c
+
+julia> a*inv(a)
+(empty word)
+
+julia> (a*b)^2
+a*b*a*b
+
+julia> commutator(a, b)
+A*B*a*b
+
+julia> x = a*b; y = inv(b)*a;
+
+julia> x*y
+a^2
+
+```
+Let's create a quotient of the free group above:
+```julia
+julia> ε = one(F);
+
+julia> G = FPGroup(F, [a^2 => ε, b^3=> ε, (a*b)^7=>ε, (a*b*a*inv(b))^6 => ε, commutator(a, c) => ε, commutator(b, c) => ε ])
+┌ Warning: Maximum number of rules (100) reached. The rewriting system may not be confluent.
+│     You may retry `knuthbendix` with a larger `maxrules` kwarg.
+└ @ KnuthBendix ~/.julia/packages/KnuthBendix/i93Np/src/kbs.jl:6
+⟨a, b, c | a^2 => (empty word), b^3 => (empty word), a*b*a*b*a*b*a*b*a*b*a*b*a*b => (empty word), a*b*a*B*a*b*a*B*a*b*a*B*a*b*a*B*a*b*a*B*a*b*a*B => (empty word), A*C*a*c => (empty word), B*C*b*c => (empty word)⟩
+
+```
+As you can see from the warning, the Knuth-Bendix procedure has not completed successfully. This means that we only are able to approximate the word problem in `G`, i.e. if the equality (`==`) of two group elements may return `false` even if group elements are equal. Let us try with a larger maximal number of rules in the underlying rewriting system.
+
+```julia
+julia> G = FPGroup(F, [a^2 => ε, b^3=> ε, (a*b)^7=>ε, (a*b*a*inv(b))^6 => ε, commutator(a, c) => ε, commutator(b, c) => ε ], maxrules=500)
+⟨a, b, c | a^2 => (empty word), b^3 => (empty word), a*b*a*b*a*b*a*b*a*b*a*b*a*b => (empty word), a*b*a*B*a*b*a*B*a*b*a*B*a*b*a*B*a*b*a*B*a*b*a*B => (empty word), A*C*a*c => (empty word), B*C*b*c => (empty word)⟩
+
+```
+This time there was no warning, i.e. Knuth-Bendix completion was successful and we may treat the equality (`==`) as true mathematical equality. Note that `G` is the direct product of `ℤ = ⟨ c ⟩` and a quotient of van Dyck `(2,3,7)`-group. Let's create a random word and reduce it as an element of `G`.
+```julia
+julia> using Random; Random.seed!(1); w = Groups.Word(rand(1:length(A), 16))
+KnuthBendix.Word{UInt16}: 4·6·1·1·1·6·5·1·5·2·3·6·2·4·2·6
+
+julia> F(w) # freely reduced w
+B*C*a^4*c*A*b*C*A*B*A*C
+
+julia> G(w) # w as an element of G
+B*a*b*a*B*a*C^2
+
+julia> F(w) # freely reduced w
+B*C*a^4*c*A*b*C*A*B*A*C
+
+julia> word(ans) # the underlying word in A
+KnuthBendix.Word{UInt8}: 4·6·1·1·1·1·5·2·3·6·2·4·2·6
+
+julia> G(w) # w as an element of G
+B*a*b*a*B*a*C^2
+
+julia> word(ans) # the underlying word in A
+KnuthBendix.Word{UInt8}: 4·1·3·1·4·1·6·6
+
+```
+As we can see the underlying words change according to where they are reduced.
+Note that a word `w` (of type `Word <: AbstractWord`) is just a sequence of numbers -- pointers to letters of an `Alphabet`. Without the alphabet `w` has no meaning.
+
+### Automorphism Groups
+
+Relatively complete is the support for the automorphisms of free groups, as given by Gersten presentation:
+```julia
+julia> saut = SpecialAutomorphismGroup(F, maxrules=100)
+┌ Warning: Maximum number of rules (100) reached. The rewriting system may not be confluent.
+│     You may retry `knuthbendix` with a larger `maxrules` kwarg.
+└ @ KnuthBendix ~/.julia/packages/KnuthBendix/i93Np/src/kbs.jl:6
+automorphism group of free group on 3 generators
+
+julia> S = gens(saut)
+12-element Vector{Automorphism{FreeGroup{Symbol},…}}:
+ ϱ₁.₂
+ ϱ₁.₃
+ ϱ₂.₁
+ ϱ₂.₃
+ ϱ₃.₁
+ ϱ₃.₂
+ λ₁.₂
+ λ₁.₃
+ λ₂.₁
+ λ₂.₃
+ λ₃.₁
+ λ₃.₂
+
+julia> x, y, z = S[1], S[12], S[6];
+
+julia> f = x*y*inv(z)
+ϱ₁.₂*λ₃.₂*ϱ₃.₂^-1
+
+julia> g = inv(z)*y*x
+ϱ₃.₂^-1*ϱ₁.₂*λ₃.₂
+
+julia> word(f), word(g)
+(KnuthBendix.Word{UInt8}: 1·12·18, KnuthBendix.Word{UInt8}: 18·1·12)
+
+```
+Even though Knuth-Bendix did not finish successfully in automorphism groups we have another ace in our sleeve to solve the word problem: evaluation.
+Lets have a look at the images of generators under those automorphisms:
+```julia
+julia> evaluate(f) # or to be more verbose...
+(a*b, b, b*c*B)
+
+julia> Groups.domain(g)
+(a, b, c)
+
+julia> Groups.evaluate!(Groups.domain(g), g)
+(a*b, b, b*c*B)
+
+```
+Since these automorphism map the standard generating set to the same new generating set, they should be considered as equal! And indeed they are:
+```julia
+julia> f == g
+true
+```
+This is what is happening behind the scenes:
+ 1. words are reduced using a rewriting system
+ 2. if resulting words are equal `true` is returned
+ 3. if they are not equal `Groups.equality_data` is computed for each argument (here: the images of generators) and the result of comparison is returned.
+
+Moreover we try to amortize the cost of computing those images. That is a hash of `equality_daata` is lazily stored  in each group element and used as needed. Essentially only if `true` is returned, but comparison of words returns `false` recomputation of images is needed (to guard against hash collisions).
+
+----
+This package was developed for computations in [1712.07167](https://arxiv.org/abs/1712.07167) and in [1812.03456](https://arxiv.org/abs/1812.03456). If you happen to use this package please cite either of them.