Skip to content

hchiam/code-explorer

Repository files navigation

Code Explorer version License: MIT

Kinda like JARVIS, but just for coding.

Setup

yarn to install dependencies.

The you can run 3 CLI tabs: (or just do yarn start - don't forget to do yarn stop when you're done)

Run database

mongod --dbpath=data

Run server

node server.js

Debug database

# https://www.tutorialspoint.com/mongodb/mongodb_query_document.htm
mongo
use takeaways
db.takeaways.insert({'some-key':'some-data'})
db.takeaways.insert({'some-key':'some-other-data'})
db.takeaways.find({'some-key':'some-data'})
db.takeaways.find()
db.takeaways.update({'some-key':'some-other-data'},{$set:{'some-key':'new data'}})
db.takeaways.find()
db.takeaways.remove({'some-key':'new data'})
db.takeaways.remove({'some-key':{$regex:/some-data/}})
db.takeaways.find()

See dependency graph

deps
# for example:
# Enter max depth:
# Enter src folder (or folder or file you'd like to inspect): public/index.js

Steps

For now, most critical are 1, 2, 5, 8, 11: (goal, nearest, ask, report/show)

  1. get goal
  2. get embeddings of props/actions
    • store these embeddings for faster access later if repeated
    • (use ANNOY to get stored index and recreate index with updated data)
  3. BONUS: get embeddings so it can get ideas of combos to try:
    • get embeddings of docs sentences that are associated with API props/actions
    • get code examples from docs API to generate example inputs
    • get embeddings of any error logs (compare to props/actions/docs embeddings)
    • get embeddings of any google suggestion sentence (compare to props/actions/docs embeddings)
  4. BONUS: check if any action combos get the goal result already (key: result, value: action combo)
  5. choose closest embeddings to goal
    • check nearness of embeddings of single actions (try shallow API first)
  6. BONUS: choose closest embeddings to error log:
    • check nearness of embeddings of single actions (try shallow API first)
    • check nearness of embeddings of combos suggested by docs/google/error logs
  7. BONUS: get different ways to make combos:
    • sequence functions (doSomething(); doSomethingElseAfter();)
    • within/chained functions, conceptually includes "=" (doSomethingAfter(doSomethingFirst()))
  8. ask/suggest before trying (for security)
  9. BONUS: try combos/single actions:
    • with delay between things tried
    • in either codepen or new sandboxed browser instances (for safety and to isolate variables)
  10. BONUS: get and store results and results frequencies for what tried:
    • how use this to check if achieved goal?
  11. report/show findings from results of actions tried (variable value + UI, which may be in a CodePen if possible)

Stepping back

This process is different from semantic code search because it also proactively checks error logs, tries combinations, and checks results (at least tries to, in a sandbox).

Other notes

ResponsiveVoice

TensorFlow.js Universal Sentence Encoder lite

https://github.com/tensorflow/tfjs-models/tree/master/universal-sentence-encoder

https://github.com/hchiam/text-similarity-test

ANNOY Node bindings

For an efficient Approximate Nearest Neighbors search of nearest embeddings (instead of checking every single embedding stored in memory every time):

https://github.com/spotify/annoy

https://github.com/jimkang/annoy-node

# for python:
pip install --user annoy

You might be interested in

https://github.com/hchiam/sourcefetch-server

https://github.com/hchiam/code-assistant