Oracle Custom Import Model Browser

This is a small standalone web app that crawls Oracle Generative AI imported-model documentation, extracts supported model rows, and exposes them through a searchable website.

It is intentionally scoped to Oracle imported-model documentation only:

• Root page: https://docs.oracle.com/en-us/iaas/Content/generative-ai/imported-models.htm
• Imported-model family pages discovered dynamically from that root page

Current Scope

The app does not search the open web. It only:

1. Fetches the Oracle imported-model landing page
2. Recursively follows Oracle imported-model subpages under the same Generative AI docs path
3. Extracts model table rows from those pages
4. Filters and displays the results in the browser

Pages that do not produce at least one extracted model row are not shown in the UI.

What It Does

• Fetches Oracle's imported-model catalog pages on the server side
• Recursively crawls imported-model subpages linked from the Oracle root page
• Extracts model families and supported imported-model rows
• Displays only pages that contain at least one extracted model
• Lets you search by family, model ID, capability, or recommended shape
• Supports wildcard queries such as qwen* and looser relative-term matches such as llma for llama
• Shows the crawled Oracle subpage list in the UI so coverage is visible
• Removes duplicate model entries and adds a Hugging Face link for each model id when available
• Lets you switch the UI between grey and dark themes
• Caches results in memory for six hours unless you hit refresh
• Requires a login before serving the application or model API
• Avoids npm and third-party Python dependencies

Search Behavior

The search box supports:

• Exact substring matching
• Prefix matching
• Wildcards such as qwen*, *vision*, or meta-*
• Lightweight fuzzy and relative-word matching such as llma -> llama

Search runs against:

• Family/page title
• Oracle docs page URL
• Table section heading
• Model ID
• Capability text
• Recommended shape

UI Behavior

The browser UI includes:

• A single search bar shared across the Oracle imported-model catalog
• A Refresh from Oracle Docs action to bypass the in-memory cache
• Theme switching between grey and dark
• A Crawled Oracle subpages section so you can verify which Oracle pages produced model results
• One card per imported-model family page
• A Hugging Face link for each model when the extracted model ID looks like a valid Hugging Face repo path

API

The app serves one JSON API endpoint:

• GET /api/models

Query parameters:

• q: optional search query
• refresh=1: force a fresh crawl instead of using the cached catalog

Example:

http://127.0.0.1:8080/api/models?q=qwen*&refresh=1

Response fields include:

• source_url
• generated_at
• page_count
• crawled_pages
• family_count
• model_count
• families
• models

Project Files

• app.py: HTTP server, Oracle crawler, extraction logic, search filtering, and JSON API
• static/index.html: page structure
• static/app.js: browser-side rendering, theme toggle, and search interaction
• static/styles.css: UI styling
• start.sh: local launcher
• requirements.txt: stable dependency entry point, currently standard-library only

Run

Preferred:

cd ./customimport
bash start.sh

Manual:

cd ./customimport
python3 app.py

The launcher:

• creates .venv if needed
• upgrades pip
• installs requirements.txt
• starts the local server

Then open:

http://127.0.0.1:8080

Optional environment variables:

PORT=8090 HOST=0.0.0.0 python3 app.py

Login credentials

The app requires a username and password. Configure them with environment variables:

APP_USER_NAME=my-user APP_PASSWORD='My-Strong-Passphrase!' python3 app.py

• APP_USER_NAME defaults to oci when unset or empty.
• APP_PASSWORD is generated as a complex 16-character value when unset or empty.
• A generated password is printed once to the application console at startup; save it from the logs before signing in.
• The GET /api/health endpoint remains unauthenticated for container health checks.

If you prefer to bootstrap the virtual environment yourself:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python app.py

Notes And Limitations

• The scraper depends on the Oracle docs HTML structure. If Oracle changes the page layout, extraction logic in app.py may need adjustment.
• Search is intentionally basic and limited to the imported-model docs content, not the open web.
• Query matching supports exact, prefix, wildcard, and lightweight fuzzy token matching.
• Duplicate rows are removed before rendering, using model ID plus capability plus recommended shape as the effective uniqueness key.
• Pages with no extracted models are intentionally filtered out of both the page list and the family card list.
• Some family pages may temporarily fail to load or parse. The UI shows those failures inline per family instead of hiding them.