llm-playground

I somewhat followed the tutorial at https://cloud.google.com/python/django/appengine to deploy to GCP App Engine.

Some aspects of this repo are based on other PHAC Django projects, e.g. the base jinja template.

Issues and fixes

Secrets

Creating the secret required some fiddling due to org-level settings.

Also the secret came out in UTF-16 and wouldn't decode. Set the SECRET_KEY manually in the gcp.env before running the following commands. (The first command is optional; use only if you want to replace the existing secret with a new value.)

gcloud secrets delete django_settings
gcloud secrets create django_settings --data-file gcp.env --locations=us-east1 --replication-policy=user-managed
gcloud secrets add-iam-policy-binding django_settings --member serviceAccount:phx-datasciencellm@appspot.gserviceaccount.com --role roles/secretmanager.secretAccessor

To update the gcp.env used in App Engine, first delete it with gcloud secrets delete django_settings then re-run the commands above.

Note that there is a gcp.env and a local .env. More about this later.

APPENGINE_URL env variable

The tutorial says to update the APPENGINE_URL in app.yaml to https://<project-name>.uc.r.appspot.com. However, if you have deployed to a region other than US central ("uc") then this will fail. Ensure to copy the actual URL from your browser's address bar. In my case, it was "ue" (US East), not "uc".

Sending emails to GC addresses

Sending login emails through SMTP (SendGrid) works for gmail addresses, but is blocked by GC M365 filtering.

A second option is to use GC Notify for email sending. But GC Notify emails also get sent to Junk Mail by M365. And the M365 link checking invalidates the magic links unless you disable a number of security features in settings.py:

# M365 "clicks" links to check them, so must be > 1
MAGICLINK_TOKEN_USES = 3
# Otherwise M365 checks will invalidate token
MAGICLINK_REQUIRE_SAME_IP = False
# As above
MAGICLINK_REQUIRE_SAME_BROWSER = False

The third method (that actually works!) is to use Power Automate to send the email. These go straight to the inbox within a few seconds. Sometimes the M365 link-checking seems to invalidate the link from this method, so best to leave the settings as above.

requirements.txt

GCP expects requirements.txt to be in the root of the django project, rather than in the repo root.

.env

The .env files should be in the repo root directory. It is not used directly for GCP, but rather the contents of the file are stored in a secret. The file is almost the same for both local dev and GCP secret.

GCP version (gcp.env)

DATABASE_URL=postgres://(get this from gcp...)
GS_BUCKET_NAME=(create a bucket for document uploads in GCP storage)
SECRET_KEY=...(something random)
GC_NOTIFY_API_KEY=...
GC_NOTIFY_TEMPLATE_ID=...
POWER_AUTOMATE_URL=https://...

# One of power_automate, django_smtp, gc_notify
MAGICLINK_METHOD=power_automate
# Prepended to the page title / top nav header
BRAND_NAME=PHAC
# Appears in new chats
BRAND_CHAT_WARNING=This system is under active development. Do not enter protected information.
# Appears next to document upload
BRAND_DOCS_WARNING=Do not upload protected information.
# Comma separated list of allowable email domains for users
ALLOWED_EMAIL_DOMAINS=phac-aspc.gc.ca

For local development, just copy gcp.env to .env and add this line:

USE_CLOUD_SQL_AUTH_PROXY=True

Before running the django server, start the cloud-sql-proxy as described in the tutorial. I needed to add the -g flag:

cloud-sql-proxy -g phx-datasciencellm:us-east1:llm-playground

TODO ideas

Codebase improvements

Refactoring LLM/embedding code to be more platform independent
General views refactoring
Django websockets with HTMX-ws, OR...
(big task) Front-end (React? Vue?) + API-only backend
- Could ultimately simplify controller (django "views")

File upload UI

Allow multiple files at once
Speed (direct upload to GCP via storage API?)
- Currently its routed through Django-Storage
Progress indicators
~~User-specific prefixes in filename to prevent clashes~~
~~Duplicate filename handling:~~
- ~~Update existing (overwrite file, update Model)~~
- ~~Prompt to overwrite vs. create new file (unique string)~~
  - ~~Keep track of original filenames?~~
  - For now, always increment filenames (do not allow update)
Warn if document already uploaded by another user & viewable
- See item (7) below

Extraction

Quality (how does the text look? Extract to markdown?)
More file types
Separate from upload function?

Indexing

Robustness (keep doing it even if user navigates away)
Speed (parallelize? if rate limits aren't the issue)
~~Extract titles~~ (task-specific model, not LLM?)
- Just using text-bison for now, MVP
Generate tags (task-specific model, not LLM?)
Make use of existing metadata when possible
~~Speed up full text retrieval~~
- ~~Store text in Document rather than DocumentChunks~~
  - Still have to store text in Chunks also due to uneven splitting
  - This is space-inefficient but faster, and plain text still small.
- ~~Save CHUNK_SIZE property in Document model~~

Retrieval

See https://www.youtube.com/watch?v=Zj5RCweUHIk for ideas
Similarity search should be augmented
- See https://xata.io/blog/postgres-full-text-search-engine
Keyword search
~~Add document titles/summaries etc. into chunks!~~
~~Increase chunk size / overlap~~
Tweak distance metrics / thresholds
- Multiply chunk distance by document summary distance? (No...)
- Try other distance metrics?

Q&A chat performance

Prompt chain (prompt, stuffing vs. map-reduce, etc)
Reduce hallucinations
- Check if there isn't enough context before answering
Improve display of sources
- i.e. expandable to show the raw results of similarity search?
~~Try chat model vs. text model~~
- Text model is better
Select individual docs or tags
Option to include full text of selected docs in context
- (as opposed to similarity search for context)
Improve performance on acronyms
- Extract acronyms from documents to a table?

Chat UI

~~Session, chat, or user settings modal~~
- ~~Select model (chat vs. code-chat model)~~
- ~~Modify system prompt~~
- ~~(Advanced) modify params (temp, tokens)~~
Share chat
Right sidebar: enable document Q&A, select docs/tags
- Or enter URL
Clean up empty chats in DB
- Prevent creation of empty chats?
- Delete empty chats in cron job?

Document management

Private vs. shared (all-users) access
~~Ability to delete files (after filename clashes resolved, see 1.)~~
Edit titles, summary, tags
- Keep track of human-edited vs. generated metadata
View shared files
- Shouldn't be able to modify tags, etc. when not owner?
"Star" files from all-users access?
Search/filter documents in document management view
- Will become necessary if there are 100+ documents
"Find similar documents"
How to integrate with official IM repositories?
- This could become an IM issue if widely used
~~Hide embeddings, similarity search from most users~~
Hide document Q&A field once implemented in Chat UI
Online sources (webpage URLs, crawl from URL, youtube link, arxiv)...

Integrated help

Show tour of application if new user
- "Don't show this again" checkbox
Draw attention to:
- "New chat" button
- Old chats appear in left sidebar
  - Delete old chats
- Settings (once implemented)
- Right sidebar (once implemented)
- My documents tab
  - Upload documents
  - Expand documents / edit metadata

French

French UI is doable; normal Django translation task
How good are the LLMs in French?
- Do French queries surface English content, and vice-versa?
System prompts, LangChain internal prompts etc. may need modification
- Bot should respond in user's selected language
Language tagging of documents / chunks / elements
- Multi-lang documents? (very common in PDFs)
- Preserve language tagging where it exists? (e.g. HTML, Word, PDF)
French document summaries
French full-text
Likely issues:
- multi-lang tags (generate both? Translate?)
- search retrieves duplicates due to multi-lang

Sharing prompts / chains

allow users to refine, share (system) prompts for different tasks
- e.g. writing editor prompt
- translation
- reformat input text as list, JSON, etc.
- extract action items with names from meeting notes
Share system prompts with link (prompt in URL?)
Prompt browser UI (SystemPrompt objects in DB or 2-message chat?)
Extensibility with different chain endpoints (in codebase)
- But common UI to call different chains
- The document search chain(s) could be the first 2?

juleskuehn / llm-playground