diff --git a/DESCRIPTION b/DESCRIPTION
index 4fd616d..42235d9 100644
--- a/DESCRIPTION
+++ b/DESCRIPTION
@@ -1,6 +1,6 @@
 Package: rgpt3
 Title: Making requests from R to the GPT-3 API
-Version: 0.2.0
+Version: 0.2.1
 Authors@R: 
     person("Bennett", "Kleinberg",  email = "bennett.kleinberg@tilburguniversity.edu", role = c("aut", "cre"))
 Description: With this package you can interact with the powerful GPT-3 models in two ways: making requests for completions (e.g., ask GPT-3 to write a novel, classify text, answer questions, etc.) and retrieving text embeddings representations (i.e., obtain a low-dimensional vector representation that allows for downstream analyses). You need to authenticate with your own Open AI API key and all requests you make count towards you token quota. For completion requests and embeddings requests, two functions each allow you to send either sinlge requests (`gpt3_single_request()` and `gpt3_single_embedding()`) or send bunch requests where the vectorised structure is used (`gpt3_requests()` and `gpt3_embeddings()`).
diff --git a/R/gpt3_embeddings.R b/R/gpt3_embeddings.R
index 513a5b8..2e7c167 100644
--- a/R/gpt3_embeddings.R
+++ b/R/gpt3_embeddings.R
@@ -62,7 +62,7 @@
 
   }
 
-  output_data = rbindlist(empty_list)
+  output_data = data.table::rbindlist(empty_list)
 
   return(output_data)
 
diff --git a/README.md b/README.md
index 2b503db..843f480 100644
--- a/README.md
+++ b/README.md
@@ -113,7 +113,7 @@
               , param_temperature = 0.4)
 ```
 
-
+Note that this completion request produced 5 (`param_n = 5`) completions for each of the three prompts, so a total of 15 completions.
 
 
 ### Embeddings
@@ -134,7 +134,7 @@
 ```{r}
 # we just take the first text here as a single text example
 my_text = travel_blog_data$gpt3[1]
-my_embeddings = gpt3_embeddings(input_var = my_text)
+my_embeddings = gpt3_single_embedding(input = my_text)
 length(my_embeddings)
 # 1024 (since the default model uses the 1024-dimensional Ada embeddings model)
 
@@ -160,7 +160,7 @@
 
 **Read this:** using GPT-3 is not free, so any interaction with the GPT-3 model(s) is counted towards you token quota. You can find details about Open AI's pricing model at [https://openai.com/api/pricing/](https://openai.com/api/pricing/).
 
-You receive a reasonable credit for free and do not need to provide any payment information for the first interactions with the model. Once you token quota nears its end, Open AI will let you know. Your usage is tracable in your Open AI account dashboard.
+You receive a reasonable credit for free and do not need to provide any payment information for the first interactions with the model. Once your token quota nears its end, Open AI will let you know. Your usage is traceable in your Open AI account dashboard.
 
 
 ## Support and contributions
diff --git a/paper/paper.bib b/paper/paper.bib
index 0ed7185..72f001a 100644
--- a/paper/paper.bib
+++ b/paper/paper.bib
@@ -110,5 +110,20 @@
 
 @article{shihadehbrilliance,
   title={Brilliance Bias in GPT-3},
-  author={Shihadeh, Juliana and Ackerman, Margareta and Troske, Ashley and Lawson, Nicole and Gonzalez, Edith}
+  author={Shihadeh, Juliana and Ackerman, Margareta and Troske, Ashley and Lawson, Nicole and Gonzalez, Edith},
+  year={2022}
+}
+
+@article{vandermaas2021,
+	title = {How much intelligence is there in artificial intelligence? A 2020 update},
+	author = {van der Maas, Han L.J. and Snoek, Lukas and Stevenson, Claire E.},
+	year = {2021},
+	month = {07},
+	date = {2021-07},
+	journal = {Intelligence},
+	pages = {101548},
+	volume = {87},
+	doi = {10.1016/j.intell.2021.101548},
+	url = {https://linkinghub.elsevier.com/retrieve/pii/S0160289621000325},
+	langid = {en}
 }
diff --git a/paper/paper.md b/paper/paper.md
index fadb079..eef9895 100644
--- a/paper/paper.md
+++ b/paper/paper.md
@@ -3,10 +3,11 @@
 tags:
 - r
 - natural language processing
-- "gpt-3"
+- gpt-3
+- language models
 - text generation
 - embeddings
-date: "4 October 2022"
+date: "23 October 2022"
 output: pdf_document
 authors:
 - name: Bennett Kleinberg
@@ -18,65 +19,134 @@
   index: 1
 - name: Department of Security and Crime Science, University College London, UK
   index: 2
+editor_options: 
+  markdown: 
+    wrap: 72
 ---
 
 # Summary
 
-The past decade has seen leap advancements in the field of Natural Language Processing (NLP, i.e., using computational methods to study human language). Of particular importance are generative language models which - among standard NLP tasks such as text classification - are able to produce text data that are often indistinguishable from human-written text. The most prominent language model is GPT-3 (short for: Generative Pre-trained Transformer 3) developed by Open AI and released to the public in 2021 [@brown2020language]. While these models offer an exciting potential for the study of human language at scale, models such as GPT-3 were also met with controversy [@bender2021dangers]. Part of the criticism stems from the opaque nature of the model and the potential biases it may hence propagate in generated text data. As a consequence, there is a need to understand the model and its limitations so researchers can use it in a responsible and informed manner. This package makes it possible to use the GPT-3 model from the R programming language, thereby opening access to this tool to the R community and enabling more researchers to use and test the powerful GPT-3 model.
-
+The past decade has seen leap advancements in the field of Natural
+Language Processing (NLP, i.e., using computational methods to study
+human language). Of particular importance are generative language models
+which - among standard NLP tasks such as text classification - are able
+to produce text data that are often indistinguishable from human-written
+text. The most prominent language model is GPT-3 (short for: Generative
+Pre-trained Transformer 3) developed by Open AI and released to the
+public in 2021 [@brown2020language]. While these models offer an
+exciting potential for the study of human language at scale, models such
+as GPT-3 were also met with controversy [@bender2021dangers]. Part of
+the criticism stems from the opaque nature of the model and the
+potential biases it may hence propagate in generated text data. As a
+consequence, there is a need to understand the model and its limitations
+so researchers can use it in a responsible and informed manner. This
+package makes it possible to use the GPT-3 model from the R programming
+language, thereby opening access to this tool to the R community and
+enabling more researchers to use, test and study the powerful GPT-3
+model.
 
 # Statement of need
 
-The GPT-3 model has pushed the boundaries the language abilities of artificially intelligent systems. Many tasks that were deemed unrealistic or too difficult for computational models are now solvable. 
+The GPT-3 model has pushed the boundaries the language abilities of
+artificially intelligent systems. Many tasks that were deemed
+unrealistic or too difficult for computational models are now deemed
+solvable [@vandermaas2021]. Especially the performances of the model on
+tasks originating from Psychology show the enormous potential of the
+GPT-3 model. For example, when asked to formulate creative use cases of
+everyday objects (e.g., a fork), the GPT-3 model produced alternative
+uses of the objects that were rated of higher utility (but lower
+originality and surprise) than creative use cases produced by human
+participants [@stevenson2022putting]. Others found that the GPT-3 model
+shows verbal behaviour similar to humans on cognitive tasks so much so
+that the model made the same intuitive mistakes that are observed in
+humans [@binz2022using]. Aside from these efforts to understand *how the
+model thinks*, others started to understand the personality that may be
+represented by the model. Asked to fill-in a standard personality
+questionnaire and a human values survey, the GPT-3 model showed a
+response pattern comparable with human samples and showed evidence of
+favouring specific values over others (e.g., self-direction \>
+conformity) [@miotto_who_2022].
 
-GPT-3 changes how we do research
-and what NLP/AI can do
-van der maas
+There is also ample evidence that the GPT-3 model produces biased
+responses (e.g., assigning attributes of brilliance more often to men
+than to women) [@shihadehbrilliance]. Both the promises and challenges
+with the GPT-3 model require that we start to understand the system
+better. Of particular relevance in the ambition to study such a black
+box language model is the "machine behaviour" [@rahwan2019machine]
+approach, which harnesses research designs from psychological and social
+science research to map out the behaviour and processes of algorithms
+(e.g. GPT-3).
 
-
-powerful tool with many promises and dangers
-bias
-bias
-bias
-
-binz
-stevenson
-miotto
-
-
-need to understand the system
- researchers have recently started to study GPT-3 in a "machine behaviour" [@rahwan2019machine] approach 
-need to use it
-need to have R access on it
-current barrier to using it
-
-
-Especially the performances of the model on tasks originating from Psychology show the enormous potential of large language models. For example, when asked to formulate creative use cases of everyday objects (e.g., a fork), the GPT-3 model produced alternative uses of the objects that were rated of higher utility but lower originality and surprise compared to creative use cases produced by human participants [@STEVENSON]. Others found that the GPT-3 model shows verbal behaviour similar to humans on cognitive tasks so much so that the model made the same intuitive mistakes that are observed in humans [@binz2022using]. Aside from these efforts to understand how the model _thinks_, others started to study the model in the same way as psychological research is studying human participants. Asked to fill-in a standard personality questionnaire, the GPT-3 model showed as response pattern comparable with human samples [@miotto_who_2022]. The same paper also showed that the model reports to hold values 
-
-
-
-mention on the API key issue
-no need to run python in the background
-
-Since the model has been released to the public under the Open AI API, the official libraries to interact with the model are limited to python and node.js and community libraries do not yet include access to the model via R. Since a large part of the social and behavioural science research community are using R, this package is intended to widen the access to the GPT-3 model direct from R.
-
-# Open research questions
-
-temperature
+Since a large part of the behavioural and social science community who
+may be best placed to conduct such research is using the R environment,
+this package - as the first R access point to the GPT-3 model - could
+help break down barriers and increase the adoption of GPT-3 research in
+that community.
 
 # Examples
 
-The `rgpt3` package allows users to interact via R with the GPT-3 API to perform the two core functionalities: i) prompting the model for text completions and ii) obtaining embeddings representations from text input.
+The `rgpt3` package allows users to interact via R with the GPT-3 API to
+perform the two core functionalities: i) requesting **text completions**
+and ii) obtaining embeddings representations from text input.
 
+## Completions
 
-requests
-(what they are)
-(how they are used)
-(why they are controversial)
+The core idea of text completions is to provide the GPT-3 model with
+prompts which it uses as context to generate a sequence of arbitrary
+length. For example, prompts may come in the form of questions (e.g.,
+'How does the US election work?'), tasks (e.g., 'Write a diary entry of
+a professional athlete:'), or open sequences that the model should
+finish (e.g., 'Maria has started a job as a').
 
-embeddings
-(brief example)
+This package handles completions in the most efficient manner from a
+data.table or data.frame object with the `gpt3_completions()` function.
+In the example, we provide the prompts from a data.frame and ask the
+function to produce 5 completions (via the `param_n` parameter) with a
+maximum token length each of 50 (`param_max_tokens`) with a sampling
+temperature of 0.8 (`param_temperature`). Full detail on all available
+function parameters is provided in the help files (e.g.,
+`?gpt3_completions`)
 
+The `output` object contains a list with two data.tables: the text
+generations and the meta information about the request made.
 
+```{r eval=F, echo=T}
+prompt_data = data.frame(prompts = c('How does the US election work?'
+                                     , 'Write a diary entry of a professional athlete: '
+                                     , 'Maria has started a job as a ')
+                         , prompt_id = 1:3)
+                         
+output = gpt3_completions(prompt_var = prompt_data$prompts
+                 , param_max_tokens = 50
+                 , param_n = 5
+                 , param_temperature = 0.8)
+```
+
+## Embeddings
+
+The second (albeit less relevant for computational social science work)
+functionality concerns text embeddings. An embedding representation of a
+document can help, for example, to calculate the similarity between two
+pieces of text. Embeddings can be derived as follows (using the
+package-provided mini `travel_blog_data` dataset):
+
+```{r eval=F, echo=T}
+data("travel_blog_data")
+
+example_data = travel_blog_data[1:5, ]
+                         
+embeddings = gpt3_embeddings(input_var = example_data$gpt3
+                             , id_var = 1:nrow(example_data))
+```
+
+# A note on API access
+
+When the GPT-3 model was announced, it was controversial whether the
+model should be made available to the public. Open AI decided to make it
+available through an API that users can access with their own API key
+that they receive when creating an account. Using the GPT-3 API is not
+free (although at the time of writing, each user is provided with a
+small amount to get started, which is sufficient for most basic research
+ideas).
 
 # References
diff --git a/paper/paper.pdf b/paper/paper.pdf
index f85ee97..d7ed8b5 100644
--- a/paper/paper.pdf
+++ b/paper/paper.pdf
Binary files differ
diff --git a/rgpt3_0.2.1.pdf b/rgpt3_0.2.1.pdf
new file mode 100644
index 0000000..41b0303
--- /dev/null
+++ b/rgpt3_0.2.1.pdf
Binary files differ