The second iteration of my personal chatbot, based on botframework-smalltalk. See it live here.
Run the train script to generate a classifier.json for each bot within the src/bots directory. Training is performed on each bot's associated knowledgebase.csv file. Each bot corresponds to a single language, such as en or fr.
To support more languages, create a new folder (it's name should correspond to the language the bot will support) in the src/bots directory, and add the new bot/language to the Bot enum in types.ts. Remember to include a knowledgebase.csv file in the new bot/language folder before running the train script.
Before testing a bot, make sure the bot's corresponding responses.json file exists and that all the possible Answer's found in the bot's associated knowledgebase.csv file are also included as keys in responses.json.
responses.json is of type Record<string, string[]>. This allows for each bot to have multiple responses for each Answer in knowledgebase.csv.
Run the start script to start the fastify server.
Then make a POST request to /{bot} (where {bot} correlates to a value in the Bot enum found in types.ts) to test the bot on a specific language. It is also possible to test the bot without specifying a language, make a post request to /, at which point tinyld is used to detect the language of the input text before forwarding it to the bot correlating to the predicted language. If no bot correlates to the predicted language, the en bot is used for inference.
In either case, the input text should be provided as plain text in the body of the POST request.
Inference is performed using each bot's generated classifier.json with a natural BayesClassifier. The classifier returns an intent corresponding to a key in responses.json.
As each value of responses.json is of type string[], a random response is selected from the string array corresponding to the predicted intent on each inference in order that only a single response is returned by each endpoint. This also has the effect that the same input text (i.e Hello!) correlating to the same intent (i.e smalltalk.greetings.hello) may yield a different response (i.e Howdy!, Good day!) on each request.