import os
import re
import warnings
import io
from pathlib import Path
from datetime import datetime
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
os.environ['TF_ENABLE_ONEDNN_OPTS'] = '0'
warnings.filterwarnings("ignore", category=DeprecationWarning)
warnings.filterwarnings("ignore", category=UserWarning)
import streamlit as st
import chromadb
from sentence_transformers import SentenceTransformer
from huggingface_hub import snapshot_download, InferenceClient
from reportlab.lib.pagesizes import A4
from reportlab.lib.styles import getSampleStyleSheet, ParagraphStyle
from reportlab.lib.units import cm
from reportlab.lib import colors
from reportlab.platypus import SimpleDocTemplate, Paragraph, Spacer, HRFlowable, Image as RLImage
from reportlab.lib.enums import TA_LEFT, TA_CENTER
# ── CONFIGURE ────────────────────────────────────────────────
HF_DB_REPO = "Daksh17440/remote_sensing_db"
DB_LOCAL_PATH = "/tmp/rs_data"
DB_PATH = "/tmp/rs_data/vector_db_complete"
MINERU_PATH = "/tmp/rs_data/mineru_output_complete"
COLLECTION_NAME = "rs_index"
HF_LLM_MODEL = "mistralai/Mistral-7B-Instruct-v0.3"
EMBED_MODEL = "all-MiniLM-L6-v2"
TOP_K = 5
L2_THRESHOLD = 1.2
# ─────────────────────────────────────────────────────────────
CHAPTER_INFO = {
"RAS_ch1": ("Remote Sensing: Models and Methods for Image Processing (Schowengerdt)", "Chapter 1 — The Nature of Remote Sensing"),
"RAS_ch2": ("Remote Sensing: Models and Methods for Image Processing (Schowengerdt)", "Chapter 2 — Optical Radiation Models"),
"RAS_ch3": ("Remote Sensing: Models and Methods for Image Processing (Schowengerdt)", "Chapter 3 — Sensor Models"),
"RAS_ch4": ("Remote Sensing: Models and Methods for Image Processing (Schowengerdt)", "Chapter 4 — Data Models"),
"RAS_ch5": ("Remote Sensing: Models and Methods for Image Processing (Schowengerdt)", "Chapter 5 — Spectral Transforms"),
"RAS_ch6": ("Remote Sensing: Models and Methods for Image Processing (Schowengerdt)", "Chapter 6 — Spatial Transforms"),
"RAS_ch7": ("Remote Sensing: Models and Methods for Image Processing (Schowengerdt)", "Chapter 7 — Correction and Calibration"),
"RAS_ch8": ("Remote Sensing: Models and Methods for Image Processing (Schowengerdt)", "Chapter 8 — Image Registration and Fusion"),
"RAS_ch9": ("Remote Sensing: Models and Methods for Image Processing (Schowengerdt)", "Chapter 9 — Thematic Classification"),
"RAS_fig": ("Remote Sensing: Models and Methods for Image Processing (Schowengerdt)", "Figures"),
"RAS_tables": ("Remote Sensing: Models and Methods for Image Processing (Schowengerdt)", "Tables"),
"RRS_ch1": ("Radar Remote Sensing (2022)", "Chapter 1 — Introduction to RADAR Remote Sensing"),
"RRS_ch2": ("Radar Remote Sensing (2022)", "Chapter 2 — Microwave Components and Devices"),
"RRS_ch3": ("Radar Remote Sensing (2022)", "Chapter 3 — Theory of Monostatic and Bistatic Radar"),
"RRS_ch4": ("Radar Remote Sensing (2022)", "Chapter 4 — Review of Microwave Fundamentals"),
"RRS_ch5": ("Radar Remote Sensing (2022)", "Chapter 5 — Comparative Flood Area Analysis"),
"RRS_ch6": ("Radar Remote Sensing (2022)", "Chapter 6 — Subsurface Feature Identification Using L-Band SAR"),
"RRS_ch7": ("Radar Remote Sensing (2022)", "Chapter 7 — Terrestrial Water Budget Through Radar Remote Sensing"),
"RRS_ch8": ("Radar Remote Sensing (2022)", "Chapter 8 — Application of SAR Remote Sensing"),
"RRS_ch9": ("Radar Remote Sensing (2022)", "Chapter 9 — Classification of Radar Data Using Bayesian Optimization"),
"RRS_ch10": ("Radar Remote Sensing (2022)", "Chapter 10 — Modeling and Simulation of SAR"),
"RRS_ch11": ("Radar Remote Sensing (2022)", "Chapter 11 — Flood Inundation Mapping from SAR"),
"RRS_ch12": ("Radar Remote Sensing (2022)", "Chapter 12 — Performance Assessment of Phased Array L-Band SAR"),
"RRS_ch13": ("Radar Remote Sensing (2022)", "Chapter 13 — Evaluation of Speckle Filtering Methods"),
"RRS_ch14": ("Radar Remote Sensing (2022)", "Chapter 14 — Emerging Techniques of Polarimetric Interferometry"),
"RRS_ch15": ("Radar Remote Sensing (2022)", "Chapter 15 — Advanced Method for Radar Remote Sensing"),
"RRS_ch16": ("Radar Remote Sensing (2022)", "Chapter 16 — Estimating Crop Biophysical Parameters"),
"RRS_ch17": ("Radar Remote Sensing (2022)", "Chapter 17 — Fuzzy Logic for Retrieval of Kidney Bean Crop"),
"RRS_ch18": ("Radar Remote Sensing (2022)", "Chapter 18 — Monitoring Tropical Peatlands Subsidence"),
"RRS_ch19": ("Radar Remote Sensing (2022)", "Chapter 19 — North American Continental Wetland Map"),
"RRS_ch20": ("Radar Remote Sensing (2022)", "Chapter 20 — Challenges in Radar Remote Sensing"),
"RRS_ch21": ("Radar Remote Sensing (2022)", "Chapter 21 — ISRO SAR Satellite Study"),
"RRS_ch22": ("Radar Remote Sensing (2022)", "Chapter 22 — Radar Remote Sensing of Soil Moisture"),
"ORS_ch1": ("Optical Remote Sensing of Land Surface (2017)", "Chapter 1 — Radiometry in the Optical Domain"),
"ORS_ch2": ("Optical Remote Sensing of Land Surface (2017)", "Chapter 2 — Multispectral Satellite Image Processing"),
"ORS_ch3": ("Optical Remote Sensing of Land Surface (2017)", "Chapter 3 — Digital Terrain Models from Optical Data"),
"ORS_ch4": ("Optical Remote Sensing of Land Surface (2017)", "Chapter 4 — Processing Hyperspectral Images"),
"ORS_ch5": ("Optical Remote Sensing of Land Surface (2017)", "Chapter 5 — Principle and Physics of the LiDAR Measurement"),
"ORS_ch6": ("Optical Remote Sensing of Land Surface (2017)", "Chapter 6 — Airborne LiDAR Data Processing"),
"ORS_ch7": ("Optical Remote Sensing of Land Surface (2017)", "Chapter 7 — Digital Terrain Models Derived from Airborne LiDAR"),
"LSR_ch1": ("Land Surface Remote Sensing in Urban and Coastal Areas (2017)", "Chapter 1 — Optical Remote Sensing in Urban Areas"),
"LSR_ch2": ("Land Surface Remote Sensing in Urban and Coastal Areas (2017)", "Chapter 2 — Urban Scene Analysis with Mobile Mapping"),
"LSR_ch3": ("Land Surface Remote Sensing in Urban and Coastal Areas (2017)", "Chapter 3 — Satellite Imagery as a Tool for Terrain Analysis"),
"LSR_ch4": ("Land Surface Remote Sensing in Urban and Coastal Areas (2017)", "Chapter 4 — Remote Sensing and Ocean"),
"LSR_ch5": ("Land Surface Remote Sensing in Urban and Coastal Areas (2017)", "Chapter 5 — LiDAR Measurements and Applications"),
"LSR_ch6": ("Land Surface Remote Sensing in Urban and Coastal Areas (2017)", "Chapter 6 — Contributions of Airborne Topographic LiDAR"),
"LSR_ch7": ("Land Surface Remote Sensing in Urban and Coastal Areas (2017)", "Chapter 7 — Mangrove Forest Dynamics Using Very High Resolution"),
"LSR_ch8": ("Land Surface Remote Sensing in Urban and Coastal Areas (2017)", "Chapter 8 — Remote Sensing Based Monitoring"),
"ARS_ch1": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 1 — A Systematic View of Remote Sensing"),
"ARS_ch2": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 2 — Geometric Processing and Positioning Techniques"),
"ARS_ch3": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 3 — Compositing, Smoothing and Gap Filling"),
"ARS_ch4": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 4 — Data Fusion"),
"ARS_ch5": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 5 — Atmospheric Correction of Optical Images"),
"ARS_ch6": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 6 — Solar Radiation"),
"ARS_ch7": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 7 — Broadband Albedo"),
"ARS_ch8": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 8 — Land Surface Temperature and Thermal Infrared"),
"ARS_ch9": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 9 — Surface Longwave Radiation Budget"),
"ARS_ch10": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 10 — Canopy Biochemical Characteristics"),
"ARS_ch11": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 11 — Leaf Area Index"),
"ARS_ch12": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 12 — Fraction of Absorbed Photosynthetically Active Radiation"),
"ARS_ch13": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 13 — Fractional Vegetation Cover"),
"ARS_ch14": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 14 — Vegetation Height and Vertical Structure"),
"ARS_ch15": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 15 — Above-ground Biomass"),
"ARS_ch16": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 16 — Vegetation Production in Terrestrial Ecosystems"),
"ARS_ch17": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 17 — Precipitation"),
"ARS_ch18": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 18 — Terrestrial Evapotranspiration"),
"ARS_ch19": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 19 — Soil Moisture Content"),
"ARS_ch20": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 20 — Snow Water Equivalence"),
"ARS_ch21": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 21 — Water Storage"),
"ARS_ch22": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 22 — High-level Land Product Integration"),
"ARS_ch23": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 23 — Production and Data Management Systems"),
"ARS_ch24": ("Advanced Remote Sensing (Liang, 2012)", "Chapter 24 — Land Cover and Land Use Changes"),
"ERS_ch1": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 1 — Introduction"),
"ERS_ch2": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 2 — Balloon-borne Radiometers"),
"ERS_ch3": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 3 — Frost Point Hygrometers"),
"ERS_ch4": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 4 — Ozonesondes"),
"ERS_ch5": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 5 — Oceanographic Buoys"),
"ERS_ch6": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 6 — Surface-based Thermal Infrared"),
"ERS_ch7": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 7 — Sun Photometers"),
"ERS_ch8": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 8 — The AirCore Atmospheric Sampler"),
"ERS_ch9": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 9 — High Altitude Aircraft Radiometers"),
"ERS_ch10": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 10 — Aircraft Dropsondes"),
"ERS_ch11": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 11 — Ship-based Cal/Val"),
"ERS_ch12": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 12 — Land-based Cal/Val"),
"ERS_ch13": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 13 — Aircraft Vertical Profile Measurements"),
"ERS_ch14": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 14 — Campaign Situational Awareness"),
"ERS_ch15": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 15 — On-orbit VIIRS Sensor Calibration"),
"ERS_ch16": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 16 — The NOAA Sounding Products"),
"ERS_ch17": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 17 — Satellite Microwave Sounding"),
"ERS_ch18": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 18 — Considerations for Thermal Sensors"),
"ERS_ch19": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 19 — Sea Surface Temperature Validation"),
"ERS_ch20": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 20 — Satellite Ocean Color"),
"ERS_ch21": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 21 — Land Surface Temperature"),
"ERS_ch22": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 22 — Heterogeneity of Smoke from Fires"),
"ERS_ch23": ("Field Measurements for Passive Environmental Remote Sensing (2022)", "Chapter 23 — Downburst Monitoring"),
"CL25_4231": ("NASA-ISRO SAR (NISAR) Mission Science Users Handbook", "NASA"),
"RADAR_INTERFEROMETRY": ("RADAR INTERFEROMETRY", "Remote Sensing and Digital Image Processing")
}
def get_citation(source_file, chapter_name_meta):
key = chapter_name_meta or Path(source_file).stem
if key in CHAPTER_INFO:
return CHAPTER_INFO[key]
stem = Path(source_file).stem
if stem in CHAPTER_INFO:
return CHAPTER_INFO[stem]
return (source_file.replace('.pdf','').replace('_',' '), "")
def find_image_in_mineru(image_path_meta, chapter_name):
"""
Find actual image file from mineru_output_complete.
image_path_meta: path stored in ChromaDB metadata (may be old local path)
chapter_name: e.g. RAS_ch1
"""
if not image_path_meta:
return None
# First try: use stored path directly (works if DB was built on same machine)
p = Path(image_path_meta)
if p.exists():
return str(p)
# Second try: extract just the filename and find it in mineru output
img_filename = Path(image_path_meta).name
if not img_filename:
return None
# Search in the chapter's mineru folder
chapter_dir = Path(MINERU_PATH) / chapter_name
if chapter_dir.exists():
matches = list(chapter_dir.rglob(img_filename))
if matches:
return str(matches[0])
# Third try: search all mineru output by filename
all_matches = list(Path(MINERU_PATH).rglob(img_filename))
if all_matches:
return str(all_matches[0])
return None
def render_answer(text):
parts = re.split(r'(\[.*?\]|\$\$.*?\$\$)', text, flags=re.DOTALL)
for part in parts:
part = part.strip()
if not part:
continue
if (part.startswith('[') and part.endswith(']')) or \
(part.startswith('$$') and part.endswith('$$')):
latex = part.strip('[]').strip('$$').strip()
try:
st.latex(latex)
except Exception:
st.markdown(part)
else:
st.markdown(part)
def generate_chat_pdf(messages):
"""Generate a PDF of the full chat history using ReportLab."""
buffer = io.BytesIO()
doc = SimpleDocTemplate(
buffer,
pagesize=A4,
rightMargin=2*cm, leftMargin=2*cm,
topMargin=2*cm, bottomMargin=2*cm
)
styles = getSampleStyleSheet()
title_style = ParagraphStyle(
'Title', parent=styles['Title'],
fontSize=18, textColor=colors.HexColor('#1F3864'),
spaceAfter=6
)
subtitle_style = ParagraphStyle(
'Subtitle', parent=styles['Normal'],
fontSize=10, textColor=colors.grey,
spaceAfter=20
)
user_label_style = ParagraphStyle(
'UserLabel', parent=styles['Normal'],
fontSize=9, textColor=colors.white,
backColor=colors.HexColor('#2E75B6'),
leftIndent=0, borderPadding=4,
spaceAfter=4, spaceBefore=12
)
user_text_style = ParagraphStyle(
'UserText', parent=styles['Normal'],
fontSize=11, textColor=colors.HexColor('#1a1a1a'),
leftIndent=10, spaceAfter=8,
backColor=colors.HexColor('#EBF3FB'),
borderPadding=6
)
assistant_label_style = ParagraphStyle(
'AssistantLabel', parent=styles['Normal'],
fontSize=9, textColor=colors.white,
backColor=colors.HexColor('#1F3864'),
leftIndent=0, borderPadding=4,
spaceAfter=4, spaceBefore=12
)
assistant_text_style = ParagraphStyle(
'AssistantText', parent=styles['Normal'],
fontSize=11, textColor=colors.HexColor('#1a1a1a'),
leftIndent=10, spaceAfter=8,
borderPadding=6
)
citation_style = ParagraphStyle(
'Citation', parent=styles['Normal'],
fontSize=9, textColor=colors.HexColor('#404040'),
leftIndent=20, spaceAfter=3,
italics=True
)
citation_header_style = ParagraphStyle(
'CitationHeader', parent=styles['Normal'],
fontSize=9, textColor=colors.HexColor('#2E75B6'),
leftIndent=10, spaceAfter=3,
fontName='Helvetica-Bold'
)
story = []
# Title
story.append(Paragraph("Remote Sensing RAG — Chat Export", title_style))
story.append(Paragraph(
f"Exported on {datetime.now().strftime('%B %d, %Y at %H:%M')}",
subtitle_style
))
story.append(HRFlowable(width="100%", thickness=2, color=colors.HexColor('#2E75B6')))
story.append(Spacer(1, 0.4*cm))
for msg in messages:
role = msg["role"]
if role == "user":
story.append(Paragraph("🧑 You", user_label_style))
# Clean text for PDF (strip markdown symbols)
clean_text = msg["content"].replace('**','').replace('*','').replace('#','')
story.append(Paragraph(clean_text, user_text_style))
elif role == "assistant":
story.append(Paragraph("🛰️ Assistant", assistant_label_style))
clean_text = msg["content"].replace('**','').replace('*','').replace('#','')
# Strip LaTeX blocks for PDF (replace with placeholder)
clean_text = re.sub(r'\[.*?\]', '[equation]', clean_text, flags=re.DOTALL)
clean_text = re.sub(r'\$\$.*?\$\$', '[equation]', clean_text, flags=re.DOTALL)
story.append(Paragraph(clean_text, assistant_text_style))
# Citations
if msg.get("citations"):
story.append(Paragraph("📖 Sources:", citation_header_style))
for book_title, chapters in msg["citations"].items():
story.append(Paragraph(f"{book_title}", citation_style))
for ch in sorted(chapters):
story.append(Paragraph(f" • {ch}", citation_style))
# Images
if msg.get("images"):
for img_data in msg["images"]:
img_path = img_data.get("path","")
if img_path and Path(img_path).exists():
try:
rl_img = RLImage(img_path, width=12*cm, height=8*cm, kind='proportional')
story.append(rl_img)
story.append(Paragraph(
f"Figure from: {img_data.get('chapter','')}",
citation_style
))
story.append(Spacer(1, 0.3*cm))
except Exception:
pass
story.append(Spacer(1, 0.2*cm))
story.append(Spacer(1, 0.5*cm))
story.append(HRFlowable(width="100%", thickness=1, color=colors.HexColor('#CCCCCC')))
story.append(Paragraph(
"Generated by Remote Sensing RAG System",
ParagraphStyle('Footer', parent=styles['Normal'],
fontSize=8, textColor=colors.grey,
alignment=TA_CENTER, spaceBefore=6)
))
doc.build(story)
buffer.seek(0)
return buffer
# ── LOAD RESOURCES ────────────────────────────────────────────
@st.cache_resource(show_spinner="Loading knowledge base...")
def load_resources():
data_path = Path(DB_LOCAL_PATH)
# Download entire repo (DB + mineru output) on first startup
if not data_path.exists() or not list(data_path.rglob("*.sqlite3")):
st.info("First startup: downloading data (~may take 2-3 min)...")
snapshot_download(
repo_id=HF_DB_REPO,
repo_type="dataset",
local_dir=DB_LOCAL_PATH,
ignore_patterns=["*.md", ".gitattributes"]
)
chroma_client = chromadb.PersistentClient(path=DB_PATH)
collection = chroma_client.get_or_create_collection(name=COLLECTION_NAME)
embed_model = SentenceTransformer(EMBED_MODEL)
hf_token = os.environ.get("HF_TOKEN", "")
llm_client = InferenceClient(
model=HF_LLM_MODEL,
token=hf_token if hf_token else None
)
return collection, embed_model, llm_client
# ── PAGE CONFIG ───────────────────────────────────────────────
st.set_page_config(
page_title="Remote Sensing RAG",
page_icon="🛰️",
layout="wide"
)
st.title("🛰️ Remote Sensing RAG")
st.caption("Retrieval-Augmented Generation over 6 Remote Sensing Textbooks")
try:
collection, embed_model, llm_client = load_resources()
except Exception as e:
st.error(f"Failed to load: {e}")
st.stop()
# ── SIDEBAR ───────────────────────────────────────────────────
with st.sidebar:
st.header("⚙️ Settings")
top_k = st.slider("Chunks to retrieve", 1, 10, TOP_K)
threshold = st.slider("L2 Distance Threshold", 0.5, 2.0, L2_THRESHOLD, step=0.05)
show_chunks = st.toggle("Show retrieved chunks", value=False)
show_images = st.toggle("Show retrieved images", value=True)
st.divider()
# PDF Export
st.header("💾 Export")
if st.button("📄 Save chat as PDF", use_container_width=True):
if not st.session_state.get("messages"):
st.warning("No messages to export yet.")
else:
with st.spinner("Generating PDF..."):
pdf_buffer = generate_chat_pdf(st.session_state.messages)
st.download_button(
label="⬇️ Download PDF",
data=pdf_buffer,
file_name=f"rs_rag_chat_{datetime.now().strftime('%Y%m%d_%H%M')}.pdf",
mime="application/pdf",
use_container_width=True
)
st.divider()
# DB Stats
st.header("📚 Indexed Books")
try:
all_meta = collection.get(include=["metadatas"])
books = {}
for meta in all_meta['metadatas']:
ch_key = meta.get('chapter_name','') or Path(meta.get('source_file','')).stem
book_title, _ = get_citation(meta.get('source_file',''), ch_key)
books[book_title] = books.get(book_title, 0) + 1
for book, count in sorted(books.items()):
st.markdown(f"**{book}**")
st.caption(f"{count} chunks")
except Exception as e:
st.error(f"DB Error: {e}")
st.divider()
if st.button("🗑️ Clear chat", use_container_width=True):
st.session_state.messages = []
st.rerun()
st.caption(f"LLM: {HF_LLM_MODEL.split('/')[-1]}")
# ── CHAT HISTORY ──────────────────────────────────────────────
if "messages" not in st.session_state:
st.session_state.messages = []
for message in st.session_state.messages:
with st.chat_message(message["role"]):
if message["role"] == "assistant":
render_answer(message["content"])
else:
st.markdown(message["content"])
if "citations" in message and message["citations"]:
with st.expander("📖 Citations"):
for book_title, chapters in message["citations"].items():
st.markdown(f"**{book_title}**")
for ch in sorted(chapters):
st.markdown(f" • {ch}")
if "images" in message and show_images and message.get("images"):
imgs = message["images"]
if imgs:
with st.expander(f"🖼️ Retrieved Images ({len(imgs)})"):
cols = st.columns(min(len(imgs), 3))
for idx, img in enumerate(imgs):
with cols[idx % 3]:
try:
st.image(img["path"], caption=img.get("chapter",""))
except Exception:
st.caption(f"Image unavailable")
if "chunks" in message and show_chunks and message.get("chunks"):
with st.expander(f"🔍 Retrieved Chunks ({len(message['chunks'])})"):
for chunk in message["chunks"]:
st.markdown(f"**{chunk['book']}** — {chunk['chapter']} | L2: {chunk['distance']}")
st.text(chunk['text'][:300] + "...")
st.divider()
# ── QUERY ─────────────────────────────────────────────────────
if query := st.chat_input("Ask a remote sensing question..."):
st.session_state.messages.append({"role": "user", "content": query})
with st.chat_message("user"):
st.markdown(query)
with st.chat_message("assistant"):
with st.spinner("Searching knowledge base..."):
query_embedding = embed_model.encode(query).tolist()
results = collection.query(
query_embeddings=[query_embedding],
n_results=top_k,
include=["documents", "metadatas", "distances"]
)
docs = results['documents'][0]
metas = results['metadatas'][0]
distances = results['distances'][0]
filtered_docs = []
filtered_metas = []
retrieved_chunks = []
retrieved_images = []
for doc, meta, dist in zip(docs, metas, distances):
if dist > threshold:
continue
ch_key = meta.get('chapter_name','') or Path(meta.get('source_file','')).stem
book_title, chapter_title = get_citation(meta.get('source_file',''), ch_key)
block_type = meta.get('block_type','text')
filtered_docs.append(doc)
filtered_metas.append(meta)
retrieved_chunks.append({
"book": book_title,
"chapter": chapter_title,
"distance": round(dist, 3),
"text": doc,
})
# Image retrieval using the smart finder
if block_type == 'image' and show_images:
img_path_meta = meta.get('image_path','')
found_path = find_image_in_mineru(img_path_meta, ch_key)
if found_path:
retrieved_images.append({
"path": found_path,
"book": book_title,
"chapter": chapter_title
})
if not filtered_docs:
answer = (
"Not enough context to answer this question from the indexed books. "
"Try rephrasing, or check if this topic is covered in your textbooks."
)
st.warning(answer)
st.session_state.messages.append({"role": "assistant", "content": answer})
else:
context = "\n\n".join(filtered_docs)
# Last 1 exchange (2 messages) for context continuity
history = ""
recent = st.session_state.messages[-2:] if len(st.session_state.messages) >= 2 \
else st.session_state.messages
for msg in recent:
role = "User" if msg["role"] == "user" else "Assistant"
history += f"{role}: {msg['content'][:500]}\n\n"
prompt = f"""[INST] You are a helpful remote sensing and geospatial analysis assistant.
Use the retrieved context to answer the question clearly and in detail.
If the answer is not in the context, say so briefly.
Retrieved Context:
{context}
Conversation History:
{history}
Current Question: {query} [/INST]"""
with st.spinner("Generating answer..."):
try:
answer = llm_client.text_generation(
prompt,
max_new_tokens=800,
temperature=0.7,
repetition_penalty=1.1,
do_sample=True,
)
render_answer(answer)
# Citations
citations = {}
for meta in filtered_metas:
ch_key = meta.get('chapter_name','') or Path(meta.get('source_file','')).stem
book_title, chapter_title = get_citation(meta.get('source_file',''), ch_key)
if book_title not in citations:
citations[book_title] = set()
if chapter_title:
citations[book_title].add(chapter_title)
with st.expander("📖 Citations"):
for book_title, chapters in citations.items():
st.markdown(f"**{book_title}**")
for ch in sorted(chapters):
st.markdown(f" • {ch}")
if retrieved_images and show_images:
with st.expander(f"🖼️ Retrieved Images ({len(retrieved_images)})"):
cols = st.columns(min(len(retrieved_images), 3))
for idx, img in enumerate(retrieved_images):
with cols[idx % 3]:
try:
st.image(img["path"], caption=img["chapter"])
except Exception:
st.caption("Image unavailable")
if retrieved_chunks and show_chunks:
with st.expander(f"🔍 Retrieved Chunks ({len(retrieved_chunks)})"):
for chunk in retrieved_chunks:
st.markdown(f"**{chunk['book']}** — {chunk['chapter']} | L2: {chunk['distance']}")
st.text(chunk['text'][:300] + "...")
st.divider()
st.session_state.messages.append({
"role": "assistant",
"content": answer,
"citations": {k: list(v) for k, v in citations.items()},
"chunks": retrieved_chunks,
"images": retrieved_images,
})
except Exception as e:
st.error(f"LLM Error: {e}")
st.info("Free HF Inference API may be rate-limited. Wait a moment and retry.")