RAG-SA / rag_hf.py
javiervzpucp's picture
up
8c421fb verified
raw
history blame
9.33 kB
# rag_interface.py (with numpy instead of faiss)
import streamlit as st
import pickle
import numpy as np
import rdflib
import torch
import datetime
import os
import requests
from rdflib import Graph as RDFGraph, Namespace
from sentence_transformers import SentenceTransformer
from dotenv import load_dotenv
# === CONFIGURATION ===
load_dotenv()
MODEL_ID = "mistralai/Mistral-7B-Instruct-v0.3"
EMBEDDING_MODEL = "intfloat/multilingual-e5-base"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
EX = Namespace("http://example.org/lang/")
st.set_page_config(
page_title="Vanishing Voices: Language Atlas",
page_icon="🌍",
layout="wide",
initial_sidebar_state="expanded"
)
# Custom CSS
st.markdown("""
<style>
.header {
color: #2c3e50;
border-bottom: 2px solid #3498db;
padding-bottom: 10px;
margin-bottom: 1.5rem;
}
.info-box {
background-color: #e8f4fc;
border-radius: 8px;
padding: 1rem;
margin-bottom: 1.5rem;
border-left: 4px solid #3498db;
}
.sidebar-section {
margin-bottom: 2rem;
}
.sidebar-title {
color: #2c3e50;
font-size: 1.1rem;
font-weight: 600;
margin-bottom: 0.5rem;
border-bottom: 1px solid #eee;
padding-bottom: 0.5rem;
}
.method-card {
background-color: #f8f9fa;
border-radius: 8px;
padding: 0.8rem;
margin-bottom: 0.8rem;
border-left: 3px solid #3498db;
}
.method-title {
font-weight: 600;
color: #3498db;
margin-bottom: 0.3rem;
}
</style>
""", unsafe_allow_html=True)
@st.cache_resource(show_spinner="Loading models and indexes...")
def load_all_components():
embedder = SentenceTransformer(EMBEDDING_MODEL, device=DEVICE)
methods = {}
for label, suffix, ttl, matrix_path in [
("Standard", "", "grafo_ttl_no_hibrido.ttl", "embed_matrix.npy"),
("Hybrid", "_hybrid", "grafo_ttl_hibrido.ttl", "embed_matrix_hybrid.npy"),
("GraphSAGE", "_hybrid_graphsage", "grafo_ttl_hibrido_graphsage.ttl", "embed_matrix_hybrid_graphsage.npy")
]:
with open(f"id_map{suffix}.pkl", "rb") as f:
id_map = pickle.load(f)
with open(f"grafo_embed{suffix}.pickle", "rb") as f:
G = pickle.load(f)
matrix = np.load(matrix_path)
rdf = RDFGraph()
rdf.parse(ttl, format="ttl")
methods[label] = (matrix, id_map, G, rdf)
return methods, embedder
methods, embedder = load_all_components()
# === CORE FUNCTIONS ===
def get_top_k(matrix, id_map, query, k):
vec = embedder.encode(f"query: {query}", convert_to_tensor=True, device=DEVICE)
vec = vec.cpu().numpy().astype("float32")
sims = np.dot(matrix, vec) / (np.linalg.norm(matrix, axis=1) * np.linalg.norm(vec) + 1e-10)
top_k_idx = np.argsort(sims)[-k:][::-1]
return [id_map[i] for i in top_k_idx]
def get_context(G, lang_id):
node = G.nodes.get(lang_id, {})
lines = [f"**Language:** {node.get('label', lang_id)}"]
if node.get("wikipedia_summary"):
lines.append(f"**Wikipedia:** {node['wikipedia_summary']}")
if node.get("wikidata_description"):
lines.append(f"**Wikidata:** {node['wikidata_description']}")
if node.get("wikidata_countries"):
lines.append(f"**Countries:** {node['wikidata_countries']}")
return "\n\n".join(lines)
def query_rdf(rdf, lang_id):
q = f"""
PREFIX ex: <http://example.org/lang/>
SELECT ?property ?value WHERE {{ ex:{lang_id} ?property ?value }}
"""
try:
return [
(str(row[0]).split("/")[-1], str(row[1]))
for row in rdf.query(q)
]
except Exception as e:
return [("error", str(e))]
def generate_response(matrix, id_map, G, rdf, user_question, k=3):
ids = get_top_k(matrix, id_map, user_question, k)
context = [get_context(G, i) for i in ids]
rdf_facts = []
for i in ids:
rdf_facts.extend([f"{p}: {v}" for p, v in query_rdf(rdf, i)])
prompt = f"""<s>[INST]
You are an expert in South American indigenous languages.
Use strictly and only the information below to answer the user question in **English**.
- Do not infer or assume facts that are not explicitly stated.
- If the answer is unknown or insufficient, say "I cannot answer with the available data."
- Limit your answer to 100 words.
### CONTEXT:
{chr(10).join(context)}
### RDF RELATIONS:
{chr(10).join(rdf_facts)}
### QUESTION:
{user_question}
Answer:
[/INST]"""
try:
res = requests.post(
f"https://api-inference.huggingface.co/models/{MODEL_ID}",
headers={"Authorization": f"Bearer {os.getenv('HF_API_TOKEN')}", "Content-Type": "application/json"},
json={"inputs": prompt}, timeout=30
)
out = res.json()
if isinstance(out, list) and "generated_text" in out[0]:
return out[0]["generated_text"].replace(prompt.strip(), "").strip(), ids, context, rdf_facts
return str(out), ids, context, rdf_facts
except Exception as e:
return str(e), ids, context, rdf_facts
# === MAIN FUNCTION ===
def main():
st.markdown("""
<h1 class='header'>Vanishing Voices: South America's Endangered Language Atlas</h1>
<div class='info-box'>
<b>Linguistic Emergency:</b> Over 40% of South America's indigenous languages face extinction.
This tool documents these cultural treasures before they disappear forever.
</div>
""", unsafe_allow_html=True)
with st.sidebar:
st.image("https://glottolog.org/static/img/glottolog_lod.png", width=180)
with st.container():
st.markdown('<div class="sidebar-title">About This Tool</div>', unsafe_allow_html=True)
st.markdown("""
<div class="method-card">
<div class="method-title">Standard Search</div>
Semantic retrieval based on text-only embeddings. Identifies languages using purely linguistic similarity from Wikipedia summaries and labels.
</div>
<div class="method-card">
<div class="method-title">Hybrid Search</div>
Combines semantic embeddings with structured data from knowledge graphs. Enriches language representation with contextual facts.
</div>
<div class="method-card">
<div class="method-title">GraphSAGE Search</div>
Leverages deep graph neural networks to learn relational patterns across languages. Captures complex cultural and genealogical connections.
</div>
""", unsafe_allow_html=True)
with st.container():
st.markdown('<div class="sidebar-title">Research Settings</div>', unsafe_allow_html=True)
k = st.slider("Languages to analyze per query", 1, 10, 3)
st.markdown("**Display Options:**")
show_ids = st.checkbox("Language IDs", value=True, key="show_ids")
show_ctx = st.checkbox("Cultural Context", value=True, key="show_ctx")
show_rdf = st.checkbox("RDF Relations", value=True, key="show_rdf")
with st.container():
st.markdown('<div class="sidebar-title">Data Sources</div>', unsafe_allow_html=True)
st.markdown("""
- Glottolog
- Wikidata
- Wikipedia
- Ethnologue
""")
query = st.text_input("Ask about indigenous languages:", "Which Amazonian languages are most at risk?")
if st.button("Analyze with All Methods") and query:
col1, col2, col3 = st.columns(3)
results = {}
for col, (label, method) in zip([col1, col2, col3], methods.items()):
with col:
st.subheader(f"{label} Analysis")
start = datetime.datetime.now()
response, lang_ids, context, rdf_data = generate_response(*method, query, k)
duration = (datetime.datetime.now() - start).total_seconds()
st.markdown(response)
st.markdown(f"⏱️ {duration:.2f}s | 🌐 {len(lang_ids)} languages")
if show_ids:
st.markdown("**Language Identifiers:**")
st.code("\n".join(lang_ids))
if show_ctx:
st.markdown("**Cultural Context:**")
st.markdown("\n\n---\n\n".join(context))
if show_rdf:
st.markdown("**RDF Knowledge:**")
st.code("\n".join(rdf_data))
results[label] = response
log = f"""
[{datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')}]
QUERY: {query}
STANDARD:
{results.get('Standard', '')}
HYBRID:
{results.get('Hybrid', '')}
GRAPH-SAGE:
{results.get('GraphSAGE', '')}
{'='*60}
"""
try:
with open("language_analysis_logs.txt", "a", encoding="utf-8") as f:
f.write(log)
except Exception as e:
st.warning(f"Failed to log: {str(e)}")
if __name__ == "__main__":
main()