├── .gitignore
├── README.md
├── main.py
├── requirements.txt
└── zero-shot-intent-classifer.chain.json


/.gitignore:
--------------------------------------------------------------------------------
1 | _venv/
2 | .env
3 | sandbox.ipynb
4 | 


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
 1 | # Zero-Shot Intent Classifier
 2 | 
 3 | This sort of thing used to be non-trivial. I hacked this together probably in like an hour. 
 4 | 
 5 | Ho boy, the times: they are a changin.
 6 | 
 7 | ## What is this?
 8 | 
 9 | This is probably going to sound archaic in a few months, but a lot of "home assistant" type devices right now use a technique called "slot filling" under the hood. An "intent" classifier is the component that figures out what the relevant slots are and 'fills' them with values, resulting in a command being emitted and arguments passed. Instead of training one bespoke: you can probably just use this directly with no or very little modification.
10 | 
11 | ## Setup
12 | 
13 | 1. `git clone <this repo>; cd <this repo>`
14 | 2. `pip install -r requirements.txt`
15 | 3. Create a file named `.env` containing one line: `OPENAI_API_KEY=...`, replacing `...` with your key.
16 | 
17 | ## Use
18 | 
19 |     $ python main.py  "becca, how would I drive from my home to SeaTac airport?"
20 |     ## {'intent': 'get_directions', 'arguments': {'start_location': 'home', 'end_location': 'SeaTac airport'}}
21 | 
22 | ## Compiled Prompt
23 | 
24 | > Act as the intent classification component of a home assistant, similar to Amazon Alexa (except your name is 'Becca', not 'Alexa').  
25 | > Common intents include: play_internet_radio, play_song_by_artist, get_weather, current_time, set_timer, remind_me  
26 | > You receive input in json format: `{"input":...}`  
27 | > You respond in json format: `{"intent":..., "arguments":{ ... }, }}`  
28 | > {"input":`{spoken_request}`}
29 | 


--------------------------------------------------------------------------------
/main.py:
--------------------------------------------------------------------------------
 1 | from dotenv import load_dotenv
 2 | from langchain.llms import OpenAI
 3 | from langchain.prompts import PromptTemplate
 4 | from langchain.chains import LLMChain
 5 | import json
 6 | 
 7 | load_dotenv()
 8 | 
 9 | INTENTS = [
10 |     "play_internet_radio", 
11 |     "play_song_by_artist",
12 |     "get_weather",
13 |     "current_time",
14 |     "set_timer",
15 |     "remind_me",
16 | ]
17 | 
18 | def build_intent_classifier(intents=INTENTS):
19 |     template = (
20 |         "Act as the intent classification component of a home assistant, similar to Amazon Alexa "
21 |         "(except your name is 'Becca', not 'Alexa').\n"
22 |         f"Common intents include: {', '.join(intents)}, ...\n"
23 |         'You receive input in json format: `{{"input":...}}`\n'
24 |         'You respond in json format: `{{"intent":..., "arguments":{{ ... }}, }}}}`\n\n'
25 |         '{{"input":{spoken_request}}}'
26 |     )
27 | 
28 |     llm = OpenAI(temperature=0.1)
29 |     prompt = PromptTemplate(
30 |         input_variables=["spoken_request"],
31 |         template=template,
32 |     )
33 |     return LLMChain(llm=llm, prompt=prompt)
34 | 
35 | def evaluate(chain, text):
36 |     response = chain.run(text)
37 |     return json.loads(response.strip())
38 | 
39 | if __name__ == '__main__':
40 |     import sys
41 | 
42 |     # e2e test
43 |     chain = build_intent_classifier()
44 |     if len(sys.argv) > 1:
45 |         text = sys.argv[1]
46 |         response = evaluate(chain, text)
47 |         print(response)
48 |     else:
49 |         text = "becca play kexp"
50 |         response = evaluate(chain, text)
51 |         print(response)
52 |         assert response == {'intent':'play_internet_radio', 'arguments':{'station':'KEXP'}}


--------------------------------------------------------------------------------
/requirements.txt:
--------------------------------------------------------------------------------
1 | langchain
2 | openai
3 | google-api-python-client
4 | python-dotenv
5 | omegaconf
6 | transformers
7 | torch
8 | 


--------------------------------------------------------------------------------
/zero-shot-intent-classifer.chain.json:
--------------------------------------------------------------------------------
 1 | {
 2 |     "memory": null,
 3 |     "verbose": false,
 4 |     "prompt": {
 5 |         "input_variables": [
 6 |             "spoken_request"
 7 |         ],
 8 |         "output_parser": null,
 9 |         "partial_variables": {},
10 |         "template": "Act as the intent classification component of a home assistant, similar to Amazon Alexa (except your name is 'Becca', not 'Alexa').\nCommon intents include: play_internet_radio, play_song_by_artist, get_weather, current_time, set_timer, remind_me, ...\nYou receive input in json format: `{{\"input\":...}}`\nYou respond in json format: `{{\"intent\":..., \"arguments\":{{ ... }}, }}}}`\n\n{{\"input\":{spoken_request}}}",
11 |         "template_format": "f-string",
12 |         "validate_template": true,
13 |         "_type": "prompt"
14 |     },
15 |     "llm": {
16 |         "model_name": "text-davinci-003",
17 |         "temperature": 0.1,
18 |         "max_tokens": 256,
19 |         "top_p": 1,
20 |         "frequency_penalty": 0,
21 |         "presence_penalty": 0,
22 |         "n": 1,
23 |         "best_of": 1,
24 |         "request_timeout": null,
25 |         "logit_bias": {},
26 |         "_type": "openai"
27 |     },
28 |     "output_key": "text",
29 |     "_type": "llm_chain"
30 | }


--------------------------------------------------------------------------------