| --- |
| license: cc-by-nc-sa-4.0 |
| language: |
| - nso |
| - en |
| task_categories: |
| - text-retrieval |
| - question-answering |
| size_categories: |
| - 1K<n<10K |
| pretty_name: EduIntel FAISS Index v1 |
| tags: |
| - sepedi |
| - sesotho-sa-leboa |
| - northern-sotho |
| - south-africa |
| - low-resource |
| - rag |
| - faiss |
| - vector-search |
| - embeddings |
| - eduintel |
| --- |
| |
| # EduIntel FAISS Index v1 |
|
|
| A **FAISS vector index** of Sepedi (Sesotho sa Leboa) text chunks for the |
| **EduIntel RAG pipeline**, built by **Sediba AI**. |
|
|
| This is **not a model** — it is a retrieval index. It stores chunked Sepedi text and their |
| embeddings so that a retriever can find relevant Sepedi content for a query, which is then |
| fed to a generative model (e.g. [Sedibaai/SedibaLM](https://huggingface.co/Sedibaai/SedibaLM)) |
| for answer generation. |
|
|
| ## Contents |
|
|
| | File | Purpose | |
| |---|---| |
| | `eduintel.index` | The FAISS index (flat or IVF vector store) | |
| | `chunks.pkl` | The chunked Sepedi text aligned to the index rows | |
| | `model.txt` | Embedding model used: **`Sediba-AI/xlmr-sepedi`** (XLM-RoBERTa-base, further pre-trained on Sepedi) | |
| | `.gitattributes` | Git LFS pointer | |
|
|
| ## Usage |
|
|
| ```python |
| import faiss, pickle |
| |
| index = faiss.read_index("eduintel.index") |
| chunks = pickle.load(open("chunks.pkl", "rb")) |
| |
| # embed your query with the SAME model listed in model.txt |
| from transformers import AutoTokenizer, AutoModel |
| tok = AutoTokenizer.from_pretrained("Sediba-AI/xlmr-sepedi") |
| mdl = AutoModel.from_pretrained("Sediba-AI/xlmr-sepedi") |
| inputs = tok("Your Sepedi query", return_tensors="pt", padding=True, truncation=True) |
| emb = mdl(**inputs).last_hidden_state.mean(dim=1).detach().numpy() |
| |
| D, I = index.search(emb, k=5) |
| retrieved = [chunks[i] for i in I[0]] |
| ``` |
|
|
| **Important:** you must embed with the same model recorded in `model.txt` |
| (`Sediba-AI/xlmr-sepedi`). A different embedding model produces vectors in an incompatible |
| space and the retrieval silently returns garbage. |
|
|
| ## What this feeds |
|
|
| This index is the **retrieval layer** of the EduIntel application — the teacher-facing |
| education intelligence surface of the TST platform. The flow is: |
|
|
| ``` |
| Sepedi query → xlmr-sepedi embed → FAISS retrieval → Sepedi chunks → SedibaLM → response |
| ``` |
|
|
| ## Licence |
|
|
| **CC BY-NC-SA 4.0** — attribution, non-commercial, share-alike. |
|
|
| This is the **interim** licence. The chunked text derives from the Sediba Sepedi training corpora |
| ([`Sediba-AI/sepedi-training-v1`](https://huggingface.co/datasets/Sediba-AI/sepedi-training-v1)), |
| which are governed by **NOODL** (Northern-Sotho Open Data Licence) pending legal review. The |
| FAISS index is a derivative work of that corpus and inherits its licence obligations. |
|
|
| **Commercial use requires a separate agreement** with Sediba AI NPC. |
|
|
| ## Limitations |
|
|
| - **No formal data-licence audit** has been completed on the underlying corpus — see the |
| `sepedi-training-v1` dataset card. |
| - **Chunk quality is not scored.** The index mixes curated, scraped and synthetic material. |
| - **Embedding model is a research artefact** (`xlmr-sepedi` has no published evaluation). |
| - **Not a knowledge base** — it is a retrieval index. It cannot answer questions by itself. |
|
|
| ## About |
|
|
| Built by **Sediba AI** — sovereign AI for South African languages, starting with Sepedi |
| (~4.7 million speakers), Mankweng, Limpopo, South Africa. |
|
|
