Fix/semantic search build time & semantic search flow after pagination refactor by vaclisinc · Pull Request #1109 · asuc-octo/berkeleytime

vaclisinc · 2026-03-13T23:35:48Z

Overview

This PR resets and simplifies the semantic search implementation after several reversals during development. The previous PR history became difficult to review, so this PR consolidates the final working version into a cleaner state.

In addition, this PR significantly speeds up Docker builds and fixes the semantic search flow after the pagination refactor.

Changes

1. Remove model pre-download from Docker build

Previously the Docker image downloaded the BAAI/bge-base-en-v1.5 model (~400MB) during build time, which made every image build slow.

This PR removes that step from the Dockerfile.

Instead, the model downloads on the first runtime start and the HuggingFace cache is persisted through a mounted volume.

Cache path:

/root/.cache/huggingface

This allows the model to persist across:

container rebuilds
pod restarts
local development runs

Local dev behavior:

docker compose up --build -d

First run → downloads the model
Later runs → loads directly from:

./data/semantic-search/model-cache/

This follows the same persistence pattern as the FAISS index.

2. Install CPU version of PyTorch

By default sentence-transformers installs the GPU build of PyTorch (~4.3GB).

Since the semantic search service only performs inference, this PR installs the CPU version of PyTorch (~1GB) to reduce image size.

3. Fix semantic search after pagination refactor

The previous implementation assumed the entire catalog was loaded on the frontend, allowing client-side filtering.

After the pagination refactor (25 courses per page), this approach no longer worked.

Semantic search is now fully handled by the backend.

Architecture

User clicks "Search with AI →"
        │
        ▼
GraphQL catalogSearch(search, semanticSearch: true)
        │
        ▼
Backend calls Python FAISS /search
        │
        ▼
FAISS returns ranked (subject, courseNumber)
        │
        ▼
Backend performs MongoDB query using those identifiers
        │
        ├─ applies user filters
        └─ retrieves course documents
        │
        ▼
Backend reorders results to match FAISS ranking
        │
        ▼
Returns CatalogResult { results, totalCount }
        │
        ▼
Frontend renders paginated results normally

Unchanged Behavior

Normal text search (Fuse.js)
Debounced search
Filters and sorting
Infinite scroll pagination
catalogSearch GraphQL response shape

Results

Semantic-search image build time improved significantly.

Before: ~20 minutes
After: ~2m 42s

≈ 86.5% faster CI build time

Review

Would appreciate a review from:

…nt variables

…ration

…y context When using git URL context with subdirectory (:apps/semantic-search), the file path must be relative to that subdirectory, not repo root. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

The label should be app.kubernetes.io/name=semantic-search, not the full deployment name Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Pre-download BAAI/bge-base-en-v1.5 model during Docker build so container doesn't need to download 420MB on every startup - Increase startupProbe to 10 minutes (from 5) for safety Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

…ables in Docker Compose

…tic search

…erms and (2) make index save in disk -> not deleted by every deployment

…nfra

- Restore deleted semantic-search module files (client.ts, controller.ts, requirements.txt) - Re-add semantic search routes to express loader - Restore ClassBrowser AI search UI components - Update fuzzy-find imports to use @repo/common - Add semantic-search to typedef validation exclusions - Restore semantic search config in packages/common Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Change import from @repo/common to @repo/common/models - Add explicit type annotation for termsWithClasses.map Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

- Re-queue failed index builds with exponential backoff (up to 10 rounds) - Retry entire startup cycle when backend isn't ready yet - Enable PVC for dev environments so indexes persist across pod restarts Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Datapuller needs this to call /refresh on the semantic search service after updating class data. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

No longer needed since we use hostPath instead of PVC. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…nfra

…nfra merge from main

ARtheboss · 2026-03-17T07:20:20Z

@vaclisinc If build times are down, should be good to merge. I'll let @PineND confirm catalog is working correctly. Will be good to merge now and monitor how heavy it's hitting our servers.

github-actions · 2026-03-18T07:12:26Z