# smartdocs/ui/components/answer_panel.py """ WHY THIS EXISTS: Consumes the SSE stream from POST /query/stream and renders: 1. Sources panel — populated immediately when 'sources' event arrives 2. Answer — streamed character by character as tokens arrive 3. Language badge — shown in answer header 4. CRAG disclosure — when answer comes from web search SSE CONSUMER DESIGN: Streamlit is synchronous. We use httpx.stream() in sync mode. st.empty() placeholders update in-place — no page rerender. Each SSE event updates a specific placeholder. Event order: sources → answer → metadata → done Sources update BEFORE the answer renders — LAW 9 compliance. THINK TAG STRIPPING: Sarvam-M exposes chain-of-thought in ... tags. These are stripped client-side before rendering. The reasoning is real but showing it confuses non-technical users. """ from __future__ import annotations import json import re import time import httpx import streamlit as st def _strip_think_tags(text: str) -> str: """ Strips Sarvam-M chain-of-thought reasoning blocks. ... content is valid reasoning — not shown to users. """ return re.sub(r".*?", "", text, flags=re.DOTALL).strip() def _render_sources(sources: list[dict], crag_triggered: bool) -> None: """Renders cited sources in an expandable panel.""" if not sources: return label = f"📚 Sources ({len(sources)})" if crag_triggered: label += " · includes web search" with st.expander(label, expanded=True): for i, source in enumerate(sources, 1): is_web = source.get("is_crag", False) icon = "🌐" if is_web else "📄" title = source.get("title", "Document") page = source.get("page_number", 0) score = source.get("reranker_score", 0.0) if is_web: url = source.get("url", "") st.markdown( f"{icon} **{title}** · [web]({url}) · relevance: `{score:.2f}`" if url else f"{icon} **{title}** · web search · relevance: `{score:.2f}`" ) else: st.markdown( f"{icon} **{title}** · page {page} · relevance: `{score:.2f}`" ) if i < len(sources): st.divider() def _render_answer_text(text: str, language: str, language_code: str) -> None: """Renders the final cleaned answer with language indicator.""" lang_flags = { "hi": "🇮🇳", "ta": "🇮🇳", "te": "🇮🇳", "kn": "🇮🇳", "ml": "🇮🇳", "mr": "🇮🇳", "en": "🇬🇧", "gu": "🇮🇳", "pa": "🇮🇳", "bn": "🇮🇳", "ur": "🇮🇳", } flag = lang_flags.get(language_code, "🌐") st.markdown( f"
" f"{flag} Answered in {language}
", unsafe_allow_html=True, ) st.markdown(text) def stream_and_render_answer( query: str, doc_id: str, doc_title: str, api_base: str, user_id: str, ) -> None: """ Calls POST /query/stream, consumes 4 SSE events, renders progressively. Rendering order (matches SSE event order): 1. Spinner shown immediately 2. 'sources' event → sources panel renders, spinner text updates 3. 'answer' event → answer renders (think tags stripped) 4. 'metadata' event → stored in session_state for cost_panel 5. 'done' event → spinner dismissed On any error: error event or exception → renders error message, never hangs. """ # Placeholders — updated in-place as events arrive spinner_placeholder = st.empty() sources_placeholder = st.empty() answer_placeholder = st.empty() spinner_placeholder.info("🔍 Searching your document...") collected_sources = [] collected_answer = "" collected_metadata = {} crag_triggered = False error_occurred = False try: with httpx.stream( "POST", f"{api_base}/query/stream", headers={ "X-User-ID": user_id, "Content-Type": "application/json", }, json={ "query": query, "doc_id": doc_id, "doc_title": doc_title, }, timeout=120.0, ) as response: if response.status_code != 200: spinner_placeholder.empty() st.error(f"Query failed: HTTP {response.status_code}") return current_event = None for line in response.iter_lines(): line = line.strip() if not line: continue if line.startswith("event:"): current_event = line[len("event:"):].strip() continue if line.startswith("data:") and current_event: raw_data = line[len("data:"):].strip() try: data = json.loads(raw_data) except json.JSONDecodeError: continue # ── sources event ───────────────────────────────────────── if current_event == "sources": collected_sources = data.get("sources", []) crag_triggered = data.get("crag_triggered", False) spinner_placeholder.info("✍️ Generating answer...") with sources_placeholder.container(): _render_sources(collected_sources, crag_triggered) # ── answer event ────────────────────────────────────────── elif current_event == "answer": raw_answer = data.get("text", "") collected_answer = _strip_think_tags(raw_answer) spinner_placeholder.empty() with answer_placeholder.container(): _render_answer_text( collected_answer, language=collected_metadata.get("language", "English"), language_code=collected_metadata.get("language_code", "en"), ) # ── metadata event ──────────────────────────────────────── elif current_event == "metadata": collected_metadata = data # Re-render answer with language info now available if collected_answer: with answer_placeholder.container(): _render_answer_text( collected_answer, language=data.get("language", "English"), language_code=data.get("language_code", "en"), ) # ── error event ─────────────────────────────────────────── elif current_event == "error": error_occurred = True spinner_placeholder.empty() error_msg = data.get("error", "Unknown error") error_code = data.get("code", "") answer_placeholder.error(f"❌ {error_msg} ({error_code})") # ── done event ──────────────────────────────────────────── elif current_event == "done": spinner_placeholder.empty() break current_event = None except httpx.TimeoutException: spinner_placeholder.empty() answer_placeholder.error( "⏱️ Query timed out after 120 seconds. " "Try a shorter or more specific question." ) error_occurred = True except httpx.ConnectError: spinner_placeholder.empty() answer_placeholder.error( "🔌 Cannot connect to SmartDocs API. " f"Is the server running at {api_base}?" ) error_occurred = True except Exception as e: spinner_placeholder.empty() answer_placeholder.error(f"❌ Unexpected error: {str(e)[:200]}") error_occurred = True # ── Persist results in session_state ────────────────────────────────────── if not error_occurred and collected_answer: st.session_state.last_answer = collected_answer st.session_state.last_sources = collected_sources st.session_state.last_metadata = collected_metadata st.session_state.last_crag = crag_triggered # Accumulate session cost cost_this_query = collected_metadata.get("cost_inr", 0.0) st.session_state.session_cost_inr = ( st.session_state.get("session_cost_inr", 0.0) + cost_this_query ) st.session_state.query_count = st.session_state.get("query_count", 0) + 1 def render_previous_answer() -> None: """ Re-renders the last answer from session_state without re-querying. Called on Streamlit reruns (e.g. after clicking a button) to preserve the answer panel state. """ if not st.session_state.get("last_answer"): return sources = st.session_state.get("last_sources", []) crag = st.session_state.get("last_crag", False) meta = st.session_state.get("last_metadata", {}) answer = st.session_state.get("last_answer", "") _render_sources(sources, crag) _render_answer_text( answer, language=meta.get("language", "English"), language_code=meta.get("language_code", "en"), )