peterpausianlian/zolai-knowledge-vectors

Zolai Knowledge Vectors

Pre-computed sentence embeddings for the Zolai-AI RAG Knowledge Brain -- a bilingual English-Zo (Tedim Chin) language preservation and learning system.

517,917 vectors from four knowledge sources, embedded with sentence-transformers/all-MiniLM-L6-v2 (384-dim).

What is Zolai?

Zolai (Tedim Zolai, ZVS 2018 orthography) is a Tibeto-Burman language spoken by the Zomi/Chin people of Myanmar and Northeast India. This dataset supports the Zolai-AI project -- a RAG-first bilingual AI toolkit to preserve and teach the language.

Dataset Overview

Source	Chunks	Description
Wiki	104,402	Grammar guides, vocabulary lists, curriculum, and learning materials from the zolai-wiki knowledge base (1,549 Markdown files)
Dictionary	114,142	Zolai-English dictionary entries (headwords, translations, parts of speech, examples) from 9 compiled dictionary sources
Parallel	248,646	Bilingual Zo-English sentence pairs aligned from the complete Zo Bible and English Bible
Bible Study	50,727	Vocabulary, grammar patterns, phrase analyses, and book summaries derived from 66 Bible books
Total	517,917	--

Schema

Each row is a JSON object:

{
  "text": "headword (pos): translation - example sentence",
  "metadata": {
    "source": "dictionary/dict_canonical_v1.jsonl",
    "source_type": "dictionary",
    "heading": "topa",
    "chunk_type": "dictionary"
  },
  "embedding": [0.0123, -0.0456],
  "embeddingModel": "sentence-transformers/all-MiniLM-L6-v2",
  "embeddingDim": 384
}

Fields

Field	Type	Description
text	string	The text content embedded (truncated to 512 chars)
metadata.source	string	Origin file path within the knowledge pipeline
metadata.source_type	string	One of: wiki, dictionary, parallel, bible
metadata.heading	string	Primary key / heading (word, verse reference, etc.)
metadata.chunk_type	string	Semantic chunk type matching source_type
embedding	list[float]	384-dimensional L2-normalized embedding vector
embeddingModel	string	HuggingFace model ID used for embedding
embeddingDim	int	Embedding dimensionality (384)

text content by source type

source_type	text format
wiki	Markdown chunk (heading + body text, up to 512 chars)
dictionary	word (pos): translations - example
parallel	English sentence --> Zo sentence (reference)
bible	word: gloss (freq: N) or pattern: example (explanation) or book: N verses, M vocab - top words

Data Sources

Source	Upstream Repo	Files
Wiki	zolai-wiki	*.md (1,549 files)
Dictionary	zolai-datasets	dict_canonical_v1.jsonl + archived sources
Parallel	zolai-datasets	zo_en_pairs_*.jsonl, bible_parallel_*.jsonl
Bible Study	zolai-datasets	*_study.jsonl (66 per-book files)

Embedding Model

• Model: sentence-transformers/all-MiniLM-L6-v2
• Dimensions: 384
• Normalization: L2-normalized (cosine similarity ready)
• Max sequence length: 512 tokens
• Build tool: zolai-core/scripts/data/build_knowledge_index.py

Usage

Load with Python

import json

vectors = []
with open("knowledge_vectors.jsonl") as f:
    for line in f:
        vectors.append(json.loads(line))

print(f"Loaded {len(vectors)} vectors")

entry = vectors[0]
print(f"Text: {entry['text'][:100]}...")
print(f"Embedding dim: {len(entry['embedding'])}")
print(f"Source: {entry['metadata']['source_type']}")

Cosine similarity search

import numpy as np
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")

query = "topa"
query_vec = model.encode([query], normalize_embeddings=True)[0]

results = []
for entry in vectors:
    sim = np.dot(query_vec, entry["embedding"])
    results.append((sim, entry))

results.sort(key=lambda x: -x[0])
for score, entry in results[:5]:
    print(f"[{score:.4f}] {entry['text'][:120]}")

With FAISS

import faiss
import json
import numpy as np

embeddings = []
with open("knowledge_vectors.jsonl") as f:
    for line in f:
        entry = json.loads(line)
        embeddings.append(entry["embedding"])

matrix = np.array(embeddings, dtype="float32")
index = faiss.IndexFlatIP(matrix.shape[1])
index.add(matrix)

print(f"FAISS index: {index.ntotal} vectors")

Language Notes

• ZVS 2018 orthography is enforced across all text content
• Key vocabulary differences vs Hakha/Falam: pasian (God) != pathian, topa (Lord) != Pathian, vantung (heaven) != van
• hiam is the universal question marker (NOT ze, which is emphatic)
• SOV word order, ergative in marker
• 95.2% dictionary coverage on Genesis after dictionary consolidation

Reproducing

git clone https://github.com/Zolai-AI/zolai-core
cd zolai-core

python scripts/data/build_knowledge_index.py \
  --data-dir /path/to/data \
  --device cuda \
  --batch-size 256

# Output: data/knowledge/knowledge_vectors.jsonl (~3.8 GB)

License

MIT -- same as all Zolai-AI repositories.

Citation

@dataset{zolai_knowledge_vectors_2026,
  title={Zolai Knowledge Vectors: English-Zo Bible + Dictionary + Wiki Embeddings},
  author={Zolai-AI},
  year={2026},
  url={https://huggingface.co/datasets/peterpausianlian/zolai-knowledge-vectors}
}

Links

• GitHub: github.com/Zolai-AI
• Website: zolai.space
• MCP Server: mcp.zolai.space/mcp
• Core Toolkit: github.com/Zolai-AI/zolai-core