Please see the slides and the 4-min demo.
This is a fork of my earlier RAG application, adapted to the RAG-PUGG project prepared for the Glia's "Hacko de Septo" hackathon (Sep 05-06, 2024).
It uses OpenAI Embeddings and Chat Completion API + Pinecone vector DB.
Note that the new code was added on top of the original application during the hackathon, so the codebase is a mix of the original logic and new flows, without any special cleaning/tests/comments etc. Besides this README section, the documentation also may be outdated and not relevant to the RAG-PUGG application. If anyone opens this code, I assume that it may be mainly to check the prompts or understand the general approach.
npm ci
-
In order to use this application, you need to populate Pinecone vector DB with documents which will be used as context for your bot when answering users' questions. Directory
scripts
contains some Python scripts which may be used for this purpose. Please seescripts/README.md
for more information. -
Copy the contents of
.env.example
file to.env
file and update the variables:
CUSTOMER_CONFIGS
is an array of objects like{ "x-customer-id": "test-customer", "x-api-key": "secret", "OPENAI_API_KEY": "secret", "OPENAI_ORG": "secret" }
:x-customer-id
is a random UUID which you need to generate once and which will serve as the ID of your customer. This value must be sent as the header keyx-customer-id
in the requests to[POST] /api/v1/complete
endpointx-api-key
- some password to protect the endpoint for this customer (headerx-api-key
)OPENAI_API_KEY
andOPENAI_ORG
- please find those at https://platform.openai.com/api-keys
POSTGRESQL_CONNECTION_STRING
- a string to connect to your Postgresql DB, e.g.postgres://postgres.bwghlfwqsbwaxnafsysc:[email protected]:6543/postgres
(example from https://supabase.com/)PINECONE_ENVIRONMENT
,PINECONE_API_KEY
,PINECONE_INDEX_NAME
- please find this values in your account on PineconeSIMILARITY_SEARCH_LIMIT
,TOP_K
- these parameters are used during similarity search. Default values can be used (0.8 and 5, correspondingly)
Note that several customers can be added to CUSTOMER_CONFIGS
, each with their own OpenAI credentials and assistant.
npm run start:local
curl --location 'localhost:3000/api/v1/complete' \
--header 'x-api-key: secret' \
--header 'x-customer-id: 74a2aeb0-6963-4eb2-b458-e62877fcc152' \
--header 'Content-Type: application/json' \
--data '{
"userId": "44e2ef2c-89a3-4428-9373-2d18d2e2113f",
"query": "Is it possible to return something I bought on clearance in the store?""
}'
Notes:
x-customer-id
andx-api-key
must coincide with the values stored in .envuserId
must be a UUID. Each user gets their own thread with the bot, in which all the messages are processed (so consider the possible effect of the previous conversation turns on the current question). Pass a newuserId
to start conversation from scratch.
Please also see https://github.com/IuriiD/ai24support-openai-assistants-api for comparison how the same task can be solved using OpenAI Assistants API.