README.md

---
title: Ask GC Library Guides
emoji: 📚
colorFrom: blue
colorTo: green
sdk: streamlit
sdk_version: 1.44.0
app_file: app.py
pinned: false
hf_oauth: false
hf_username: ''
hf_token: ''
hf_private: false
hf_space_id: ''
hf_disable_embedding: false
hf_disable_inference: false
hf_disable_sharing: false
hf_disable_suggestion: false
hf_suggested_questions: []
hf_suggested_themes: []
hf_suggested_examples: []
hf_suggested_datasets: []
hf_suggested_models: []
hf_suggested_tasks: []
hf_suggested_libraries: []
hf_suggested_metrics: []
hf_suggested_visualizers: []
hf_suggested_widgets: []
hf_suggested_co2: []
hf_suggested_pipeline_tags: []
hf_suggested_tags: []
hf_suggested_configs: []
hf_suggested_args: []
hf_suggested_kwargs: []
hf_suggested_env: {}
hf_suggested_requirements: []
hf_suggested_setup: ''
hf_suggested_dockerfile: ''
hf_suggested_app_file: ''
hf_suggested_sdk: ''
hf_suggested_sdk_version: ''
hf_suggested_python_version: ''
hf_suggested_base_image: ''
hf_suggested_entrypoint: ''
hf_suggested_cmd: ''
hf_suggested_workdir: ''
hf_suggested_expose: []
hf_suggested_volumes: []
hf_suggested_ports: []
hf_suggested_networks: []
hf_suggested_depends_on: []
hf_suggested_links: []
hf_suggested_extra_hosts: []
hf_suggested_dns: []
hf_suggested_dns_search: []
hf_suggested_cap_add: []
hf_suggested_cap_drop: []
hf_suggested_cgroup_parent: ''
hf_suggested_devices: []
hf_suggested_device_requests: []
hf_suggested_device_cgroup_rules: []
hf_suggested_dns_opt: []
hf_suggested_domainname: ''
hf_suggested_entrypoint_args: []
hf_suggested_env_file: []
hf_suggested_expose_ports: []
hf_suggested_external_links: []
hf_suggested_extra_hosts_list: []
hf_suggested_healthcheck: {}
hf_suggested_hostname: ''
hf_suggested_init: false
hf_suggested_ipc: ''
hf_suggested_labels: {}
hf_suggested_links_list: []
hf_suggested_logging: {}
hf_suggested_mac_address: ''
hf_suggested_network_mode: ''
hf_suggested_networks_list: []
hf_suggested_pid: ''
hf_suggested_ports_list: []
hf_suggested_privileged: false
hf_suggested_read_only: false
hf_suggested_restart: ''
hf_suggested_security_opt: []
hf_suggested_shm_size: ''
hf_suggested_stdin_open: false
hf_suggested_stop_grace_period: ''
hf_suggested_stop_signal: ''
hf_suggested_sysctls: {}
hf_suggested_tmpfs: []
hf_suggested_tty: false
hf_suggested_ulimits: {}
hf_suggested_user: ''
hf_suggested_userns_mode: ''
hf_suggested_volumes_from: []
hf_suggested_volumes_list: []
hf_suggested_working_dir: ''
---

# Ask GC Library Guides (RAG Demo)

This Space demonstrates a Retrieval-Augmented Generation (RAG) application built with Streamlit. It allows users to ask questions about the CUNY Graduate Center library guides.

**How it works:**

1.  **Data Source:** Pre-computed embeddings (`BAAI/bge-m3`), documents, and metadata loaded from the Hugging Face Dataset `Zwounds/Libguides_Embeddings` (originally sourced from `extracted_content.jsonl`).
2.  **Database Initialization:** On startup, the application downloads the dataset and loads the data into an in-memory ChromaDB collection stored in a temporary directory. This avoids slow re-embedding on every startup.
3.  **Query Processing:**
    *   User queries are optionally expanded using the generation model (`google/gemma-3-27b-it` via HF API).
    *   Queries are embedded using the local `BAAI/bge-m3` model (loaded into the Space).
    *   ChromaDB performs a similarity search using the query embedding against the pre-computed document embeddings.
4.  **Generation:** The relevant chunks and the original query are passed to the `google/gemma-3-27b-it` model via the Hugging Face Inference API to generate a final answer.

**Configuration:**

*   **Embedding:** Pre-computed `BAAI/bge-m3` embeddings loaded from HF Dataset `Zwounds/Libguides_Embeddings`. Query embedding uses local `BAAI/bge-m3`.
*   **Generation Model:** `google/gemma-3-27b-it` (via HF Inference API).
*   **Requires Secret:** A Hugging Face User Access Token must be added as a Space Secret named `HF_TOKEN`.

**Note:** Startup involves downloading the dataset and loading it into the ChromaDB collection, which is much faster than re-embedding all documents.