diff --git a/docs/source/service.rst b/docs/source/service.rst
index 8071403..40bb62a 100644
--- a/docs/source/service.rst
+++ b/docs/source/service.rst
@@ -41,3 +41,5 @@ Specifically, this approach iteratively identifies the most relevant features, f
 On the other hand, this methodology does not provide asymptotic guarantees on the identification of the correct underlying causal graph, but it only addresses causality in the sense of the classical Granger definition, i.e., a feature is causally relevant for a target variable if it improves the regression performance w.r.t. its autoregressive component and the other features.
 
 More specifically, the TEFS methodology can be applied in a forward and in a backward manner, with the first that may be preferred for computational and efficiency reasons. Additionally, as for the PCMCI, the choice of the time lag to consider both for the features and the target have an impact on the final output, since they balance the amount of past timestaps that the user considers to be relevant for the actual target.
+
+The final output of this process is composed of seven files. Three .pdf files contain the images that summarize the results. The first two files show the selected causal features and the related regression scores for the PCMCI and the TEFS, with the different configurations selected in the inputs. Then, the third file shows an image related to the evolution of the regression score by iteratively adding one feature to the set of the selected causal features, which can be seen as a wrapper variation of the TEFS. Finally, four pickle files compose the remaining set of outputs of this process, containing all the relevant details in terms of selected features and regression scores for the TEFS, the PCMCI, the TEFS variation as a wrapper approach, and a set of baselines.