Fix local LLM development setup: env loading, vector dimensions, and gitignore

[androemeda]

Closes #43

Summary

Fixes issues that prevent the backend from running correctly when using the local LLM setup (Ollama) as described in SETUP.md.

---

Changes

1. backend/app/core/config.py

Problem

Settings(BaseSettings) does not load the .env file when running the backend directly with:

poetry run uvicorn app.main:app --reload --host 0.0.0.0 --port 8000

It only reads from OS environment variables, which works with Docker Compose (via env_file:) but fails during local development.

Fix

Added:

model_config = SettingsConfigDict(env_file=".env")

so Pydantic automatically loads variables from the .env file.

---

2. backend/app/services/knowledge_base_service.py

Problem

The Qdrant collection is hardcoded with:

vector_size = 1536

This matches Azure OpenAI’s text-embedding-ada-002, but nomic-embed-text (recommended in SETUP.md for local use) produces 768-dimensional vectors.

This causes the following error when uploading PDFs:

Vector dimension error: expected dim: 1536, got 768

Fix

Set vector_size dynamically based on the PROVIDER setting:

• 768 → local (nomic-embed-text)
• 1536 → azure (text-embedding-ada-002)

---

3. .gitignore

Problem

The qdrant_storage/ directory (created when running Qdrant with a local volume mount as described in SETUP.md) is not ignored and may accidentally be committed.

Fix

Added:

qdrant_storage/

to .gitignore.

---

Before

• Backend fails to start when running locally with PROVIDER=local
• .env variables are not loaded when running with uvicorn
• PDF upload fails with vector dimension mismatch (1536 vs 768)
• qdrant_storage/ files can accidentally be committed

---

After

• Backend starts correctly using variables from .env
• Local Ollama setup works as documented in SETUP.md
• PDF uploads succeed with nomic-embed-text embeddings
• Qdrant storage files are properly ignored

---

Testing

• Verified backend starts successfully with:

poetry run uvicorn app.main:app --reload --host 0.0.0.0 --port 8000

• Verified /kb/upload-pdf works with nomic-embed-text embeddings (768 dimensions)
• Verified /biomodel endpoint returns data correctly

---

Notes for Reviewers

• The vector dimension logic assumes:

  • 768 for local embeddings (nomic-embed-text)
  • 1536 for Azure embeddings (text-embedding-ada-002)

• If additional embedding providers are added in the future, this logic may need to be extended.

---

Happy to make any changes if needed. Thanks for reviewing!