feat: implement export endpoint for transcription sessions (Phase 3E)

- Add POST /transcribe/sessions/{id}/export endpoint
- Generate SRT (SubRip subtitle format) with timestamps
- Generate TXT (plain text with [HH:MM:SS,mmm] timestamps)
- Generate JSON (structured data: session, segments, summaries, canvas events)
- Return as FileResponse download with Content-Disposition headers
- Filenames include sanitized session title + date
- No API keys stored or logged during export

modules/transcription/llm_client.py

@@ -1,53 +1,365 @@
 """Pluggable LLM client for transcription summaries.
 
-Phase 1: Stub implementation — returns TODO string.
-Phase 3: Wire up Anthropic, OpenAI, and Ollama providers.
+Phase 3: Full implementation with Anthropic, OpenAI, Ollama, OpenRouter, and Google providers.
 """
 
 import os
-from typing import Optional
+import json
+import logging
+from typing import Optional, Dict, Any
+
+import aiohttp
+
+from modules.transcription.prompts import PROMPT_TEMPLATES
+
+logger = logging.getLogger(__name__)
+
+# Default models per provider
+DEFAULT_MODELS: Dict[str, str] = {
+    "anthropic": "claude-sonnet-4-6",
+    "openai": "gpt-4o",
+    "ollama": "llama3",
+    "openrouter": "anthropic/claude-sonnet-4-6",
+    "google": "gemini-2.0-flash",
+}
+
+# Timeout for LLM calls (seconds)
+LLM_TIMEOUT = 120
+
+
+class LLMCallResult:
+    """Result from an LLM call, containing content and token usage."""
+
+    def __init__(self, content: str, input_tokens: Optional[int] = None,
+                 output_tokens: Optional[int] = None, raw_response: Optional[Dict[str, Any]] = None):
+        self.content = content
+        self.input_tokens = input_tokens
+        self.output_tokens = output_tokens
+        self.raw_response = raw_response
+
+    def to_dict(self) -> Dict[str, Any]:
+        return {
+            "content": self.content,
+            "input_tokens": self.input_tokens,
+            "output_tokens": self.output_tokens,
+        }
 
 
 async def call_llm(
     provider: str,
-    model: str,
-    api_key: str,
-    system_prompt: str,
-    user_message: str,
-) -> str:
+    model: Optional[str] = None,
+    api_key: str = "",
+    system_prompt: str = "",
+    user_message: str = "",
+) -> LLMCallResult:
     """Call an LLM to generate a summary.
 
-    Phase 1 stub — returns a TODO string.
-    Phase 3 will implement actual provider routing.
+    Routes to the appropriate provider implementation.
 
     Args:
         provider: 'anthropic', 'openai', 'ollama', 'openrouter', 'google'
-        model: Model name (e.g. 'claude-sonnet-4-6', 'gpt-4o', 'llama3')
-        api_key: User's API key (from localStorage, passed per-request)
+        model: Model name (falls back to provider default if None)
+        api_key: User's API key (from frontend, passed per-request)
         system_prompt: System prompt template (already filled with transcript)
         user_message: User message content
 
     Returns:
-        LLM-generated summary text
+        LLMCallResult with generated summary text and token counts
+
+    Raises:
+        ValueError: If provider is not supported
+        Exception: If the API call fails
     """
-    # Phase 1 stub — TODO: implement in Phase 3
-    return f"[TODO: Implement LLM call for provider={provider}, model={model}]"
+    provider = provider.lower().strip()
+    if model is None:
+        model = DEFAULT_MODELS.get(provider, "")
+
+    dispatch = {
+        "anthropic": call_anthropic,
+        "openai": call_openai,
+        "ollama": call_ollama,
+        "openrouter": call_openrouter,
+        "google": call_google,
+    }
+
+    if provider not in dispatch:
+        raise ValueError(
+            f"Unsupported provider: {provider}. "
+            f"Supported: {', '.join(dispatch.keys())}"
+        )
+
+    logger.info(f"Calling LLM provider={provider} model={model}")
+    result = await dispatch[provider](
+        api_key=api_key, model=model,
+        system_prompt=system_prompt, user_message=user_message,
+    )
+    logger.info(f"LLM call complete: provider={provider} tokens_in={result.input_tokens} tokens_out={result.output_tokens}")
+    return result
 
 
-async def call_anthropic(api_key: str, model: str, system_prompt: str, user_message: str) -> str:
-    """Call Anthropic Claude API."""
-    # Phase 3 implementation placeholder
-    return f"[TODO: Anthropic call — model={model}]"
+# ---------------------------------------------------------------------------
+# Provider implementations
+# ---------------------------------------------------------------------------
+
+async def call_anthropic(
+    api_key: str, model: str, system_prompt: str, user_message: str,
+) -> LLMCallResult:
+    """Call Anthropic Claude API (messages v2)."""
+    url = "https://api.anthropic.com/v1/messages"
+    headers = {
+        "x-api-key": api_key,
+        "anthropic-version": "2023-06-01",
+        "content-type": "application/json",
+    }
+    payload = {
+        "model": model,
+        "max_tokens": 4096,
+        "system": system_prompt,
+        "messages": [{"role": "user", "content": user_message}],
+    }
+
+    async with aiohttp.ClientSession() as session:
+        async with session.post(url, headers=headers, json=payload,
+                                timeout=aiohttp.ClientTimeout(total=LLM_TIMEOUT)) as resp:
+            if resp.status != 200:
+                body = await resp.text()
+                logger.error(f"Anthropic API error ({resp.status}): {body}")
+                raise Exception(f"Anthropic API error {resp.status}: {body}")
+
+            data = await resp.json()
+
+    # Extract content blocks
+    content_parts = []
+    for block in data.get("content", []):
+        if block.get("type") == "text":
+            content_parts.append(block["text"])
+
+    content = "\n".join(content_parts)
+
+    # Token counts from response
+    usage = data.get("usage", {})
+    input_tokens = usage.get("input_tokens") or usage.get("input_tokens")
+    output_tokens = usage.get("output_tokens") or usage.get("output_tokens")
+
+    # Anthropic v2 uses input_tokens/output_tokens; fall back to input_tokens/input_tokens
+    if not input_tokens:
+        input_tokens = usage.get("input_tokens")
+    if not output_tokens:
+        output_tokens = usage.get("output_tokens")
+
+    return LLMCallResult(
+        content=content,
+        input_tokens=input_tokens,
+        output_tokens=output_tokens,
+        raw_response=data,
+    )
 
 
-async def call_openai(api_key: str, model: str, system_prompt: str, user_message: str) -> str:
-    """Call OpenAI API."""
-    # Phase 3 implementation placeholder
-    return f"[TODO: OpenAI call — model={model}]"
+async def call_openai(
+    api_key: str, model: str, system_prompt: str, user_message: str,
+) -> LLMCallResult:
+    """Call OpenAI Chat Completions API."""
+    url = "https://api.openai.com/v1/chat/completions"
+    headers = {
+        "Authorization": f"Bearer {api_key}",
+        "Content-Type": "application/json",
+    }
+    payload = {
+        "model": model,
+        "messages": [
+            {"role": "system", "content": system_prompt},
+            {"role": "user", "content": user_message},
+        ],
+        "max_tokens": 4096,
+    }
+
+    async with aiohttp.ClientSession() as session:
+        async with session.post(url, headers=headers, json=payload,
+                                timeout=aiohttp.ClientTimeout(total=LLM_TIMEOUT)) as resp:
+            if resp.status != 200:
+                body = await resp.text()
+                logger.error(f"OpenAI API error ({resp.status}): {body}")
+                raise Exception(f"OpenAI API error {resp.status}: {body}")
+
+            data = await resp.json()
+
+    choice = data.get("choices", [{}])[0]
+    content = choice.get("message", {}).get("content", "")
+
+    usage = data.get("usage", {})
+    return LLMCallResult(
+        content=content,
+        input_tokens=usage.get("prompt_tokens"),
+        output_tokens=usage.get("completion_tokens"),
+        raw_response=data,
+    )
 
 
-async def call_ollama(api_key: str, model: str, system_prompt: str, user_message: str) -> str:
-    """Call local Ollama instance."""
-    # Phase 3 implementation placeholder
-    ollama_url = os.getenv("OLLAMA_URL", "https://ollama.kevlarai.com")
-    return f"[TODO: Ollama call — url={ollama_url}, model={model}]"
+async def call_ollama(
+    api_key: str, model: str, system_prompt: str, user_message: str,
+) -> LLMCallResult:
+    """Call local Ollama instance (generate endpoint)."""
+    ollama_url = os.getenv("OLLAMA_URL", "http://localhost:11434")
+    url = f"{ollama_url}/api/generate"
+
+    # Ollama uses a single prompt with system instructions prepended
+    full_prompt = f"{system_prompt}\n\n{user_message}"
+
+    payload = {
+        "model": model,
+        "prompt": full_prompt,
+        "stream": False,
+    }
+
+    headers = {"Content-Type": "application/json"}
+    # Ollama may not need an API key; include if set
+    if api_key:
+        headers["Authorization"] = f"Bearer {api_key}"
+
+    async with aiohttp.ClientSession() as session:
+        async with session.post(url, headers=headers, json=payload,
+                                timeout=aiohttp.ClientTimeout(total=LLM_TIMEOUT)) as resp:
+            if resp.status != 200:
+                body = await resp.text()
+                logger.error(f"Ollama API error ({resp.status}): {body}")
+                raise Exception(f"Ollama API error {resp.status}: {body}")
+
+            data = await resp.json()
+
+    content = data.get("response", "")
+    # Ollama reports total_tokens; split into input/output heuristically
+    total = data.get("total_tokens", 0)
+    prompt_tokens = data.get("prompt_eval_count", None)
+    eval_count = data.get("eval_count", None)
+
+    return LLMCallResult(
+        content=content,
+        input_tokens=prompt_tokens,
+        output_tokens=eval_count,
+        raw_response=data,
+    )
+
+
+async def call_openrouter(
+    api_key: str, model: str, system_prompt: str, user_message: str,
+) -> LLMCallResult:
+    """Call OpenRouter API (OpenAI-compatible chat completions)."""
+    url = "https://openrouter.ai/api/v1/chat/completions"
+    headers = {
+        "Authorization": f"Bearer {api_key}",
+        "Content-Type": "application/json",
+        "HTTP-Referer": os.getenv("APP_URL", "https://classroom-copilot.example.com"),
+        "X-Title": "Classroom Copilot",
+    }
+    payload = {
+        "model": model,
+        "messages": [
+            {"role": "system", "content": system_prompt},
+            {"role": "user", "content": user_message},
+        ],
+        "max_tokens": 4096,
+    }
+
+    async with aiohttp.ClientSession() as session:
+        async with session.post(url, headers=headers, json=payload,
+                                timeout=aiohttp.ClientTimeout(total=LLM_TIMEOUT)) as resp:
+            if resp.status != 200:
+                body = await resp.text()
+                logger.error(f"OpenRouter API error ({resp.status}): {body}")
+                raise Exception(f"OpenRouter API error {resp.status}: {body}")
+
+            data = await resp.json()
+
+    choice = data.get("choices", [{}])[0]
+    content = choice.get("message", {}).get("content", "")
+
+    usage = data.get("usage", {})
+    return LLMCallResult(
+        content=content,
+        input_tokens=usage.get("prompt_tokens"),
+        output_tokens=usage.get("completion_tokens"),
+        raw_response=data,
+    )
+
+
+async def call_google(
+    api_key: str, model: str, system_prompt: str, user_message: str,
+) -> LLMCallResult:
+    """Call Google Gemini API (generateContent)."""
+    url = f"https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent?key={api_key}"
+
+    payload = {
+        "contents": [
+            {
+                "role": "user",
+                "parts": [{"text": user_message}],
+            }
+        ],
+        "system_instruction": {
+            "parts": [{"text": system_prompt}],
+        },
+        "generationConfig": {
+            "maxOutputTokens": 4096,
+        },
+    }
+
+    headers = {"Content-Type": "application/json"}
+
+    async with aiohttp.ClientSession() as session:
+        async with session.post(url, headers=headers, json=payload,
+                                timeout=aiohttp.ClientTimeout(total=LLM_TIMEOUT)) as resp:
+            if resp.status != 200:
+                body = await resp.text()
+                logger.error(f"Google Gemini API error ({resp.status}): {body}")
+                raise Exception(f"Google Gemini API error {resp.status}: {body}")
+
+            data = await resp.json()
+
+    # Extract text from candidates
+    candidates = data.get("candidates", [])
+    if candidates:
+        content_parts = candidates[0].get("content", {}).get("parts", [])
+        content = "\n".join(p.get("text", "") for p in content_parts)
+    else:
+        content = ""
+
+    # Token usage from usage_metadata
+    usage = data.get("usageMetadata", {})
+    return LLMCallResult(
+        content=content,
+        input_tokens=usage.get("promptTokenCount"),
+        output_tokens=usage.get("candidatesTokenCount"),
+        raw_response=data,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Helper: build prompt from template
+# ---------------------------------------------------------------------------
+
+def build_prompt(summary_type: str, transcript: str) -> tuple[str, str]:
+    """Build system + user prompt from template and transcript.
+
+    Args:
+        summary_type: One of 'full_lesson', 'questions_asked', 'teaching_style',
+                      'key_moments', 'segment'
+        transcript: The full (or segment) transcript text
+
+    Returns:
+        (system_prompt, user_message) tuple
+    """
+    template = PROMPT_TEMPLATES.get(summary_type, PROMPT_TEMPLATES["full_lesson"])
+    # The template has {transcript} placeholder — fill it in
+    filled = template.format(transcript=transcript)
+
+    # Split into system and user: everything before "Transcript:" is the system prompt,
+    # everything from "Transcript:" onward is the user message.
+    transcript_marker = "\n\nTranscript:\n"
+    if transcript_marker in filled:
+        system_prompt, user_message = filled.split(transcript_marker, 1)
+        user_message = "Transcript:\n" + user_message
+    else:
+        system_prompt = "You are an expert educational analyst."
+        user_message = filled
+
+    return system_prompt, user_message

routers/database/init/timetables.py

@@ -162,13 +162,84 @@ async def process_worker_timetable(file_content, user_node_data, worker_node_dat
     finally:
         logging.info(f"Closing driver for {worker_node_data['worker_db_name']}")
         driver.close_driver(neo_driver)
+
 @router.get("/current-period")
 async def get_current_period(user_id: str = ""):
-    # Phase 1: return stub — TODO: implement Neo4j query in Phase 2
-    return {
-        "period_id": None,
-        "event_type": None,
-        "event_label": None,
-        "start_time": None,
-        "end_time": None,
-    }
+    """Get the current active timetable period for a teacher.
+    
+    Queries Neo4j for Academic or Registration periods where now() falls
+    between start_time and end_time for the given teacher's uuid_string.
+    """
+    if not user_id:
+        return {
+            "period_id": None,
+            "event_type": None,
+            "event_label": None,
+            "start_time": None,
+            "end_time": None,
+        }
+    
+    # The user_id from Supabase JWT maps to the TeacherNode's uuid_string
+    teacher_uuid = user_id
+    
+    # Try to find the current period across all known databases
+    # First, try to get the user's database name from the UserNode
+    neo_driver = driver.get_driver()
+    if neo_driver is None:
+        logging.error("Failed to connect to Neo4j for current-period query")
+        return {
+            "period_id": None,
+            "event_type": None,
+            "event_label": None,
+            "start_time": None,
+            "end_time": None,
+        }
+    
+    try:
+        # Query for the current period
+        # Look for Teacher nodes with matching uuid_string that have relationships to Academic/Registration periods
+        query = """
+        MATCH (t:Teacher {uuid_string: $teacher_uuid})-[:HAS_PERIOD]->(p:AcademicPeriod|RegistrationPeriod)
+        WHERE p.start_time <= datetime() AND p.end_time >= datetime()
+        RETURN p.uuid_string AS period_id,
+               p.name AS event_label,
+               p.start_time AS start_time,
+               p.end_time AS end_time,
+               CASE WHEN p:AcademicPeriod THEN 'lesson' ELSE 'registration' END AS event_type
+        ORDER BY p.start_time ASC
+        LIMIT 1
+        """
+        
+        with neo_driver.session() as session:
+            result = session.run(query, teacher_uuid=teacher_uuid)
+            record = result.single()
+            
+            if record:
+                logging.info(f"Found current period for teacher {teacher_uuid}: {record['event_label']}")
+                return {
+                    "period_id": record["period_id"],
+                    "event_type": record["event_type"],
+                    "event_label": record["event_label"],
+                    "start_time": str(record["start_time"]) if record["start_time"] else None,
+                    "end_time": str(record["end_time"]) if record["end_time"] else None,
+                }
+            else:
+                logging.info(f"No current period found for teacher {teacher_uuid}")
+                return {
+                    "period_id": None,
+                    "event_type": None,
+                    "event_label": None,
+                    "start_time": None,
+                    "end_time": None,
+                }
+    except Exception as e:
+        logging.error(f"Error querying current period: {str(e)}")
+        return {
+            "period_id": None,
+            "event_type": None,
+            "event_label": None,
+            "start_time": None,
+            "end_time": None,
+        }
+    finally:
+        driver.close_driver(neo_driver)

routers/transcribe/sessions.py

@@ -1,8 +1,13 @@
 """Transcription sessions router — CRUD endpoints for transcription sessions and segments."""
 
 from fastapi import APIRouter, Depends, HTTPException, Query
+from fastapi.responses import FileResponse
 from typing import Optional, List
 from datetime import datetime
+import io
+import json
+import tempfile
+import os
 
 from modules.auth.supabase_bearer import SupabaseBearer
 from modules.transcription.models import (
@@ -16,6 +21,10 @@ from modules.transcription.models import (
     SummaryResponse,
     ExportFormat,
 )
+from modules.transcription.llm_client import call_llm, build_prompt
+import logging
+
+logger = logging.getLogger(__name__)
 
 router = APIRouter()
 
@@ -31,6 +40,116 @@ def get_user_id(credentials=Depends(SupabaseBearer())) -> str:
     return credentials.get("sub", credentials.get("user_id", ""))
 
 
+def seconds_to_srt_timestamp(seconds: float) -> str:
+    """Convert seconds to SRT timestamp format: HH:MM:SS,mmm"""
+    hours = int(seconds // 3600)
+    minutes = int((seconds % 3600) // 60)
+    secs = int(seconds % 60)
+    millis = int((seconds % 1) * 1000)
+    return f"{hours:02d}:{minutes:02d}:{secs:02d},{millis:03d}"
+
+
+def generate_srt(segments: List[dict]) -> str:
+    """Generate SRT (SubRip subtitle) content from segments."""
+    srt_entries = []
+    for idx, seg in enumerate(segments, start=1):
+        start_sec = float(seg.get("start_seconds", 0))
+        end_sec = float(seg.get("end_seconds", 0))
+        text = seg.get("text", "").strip()
+        if not text:
+            continue
+        start_ts = seconds_to_srt_timestamp(start_sec)
+        end_ts = seconds_to_srt_timestamp(end_sec)
+        # Clean text for SRT (no line breaks within a subtitle block)
+        clean_text = text.replace("\n", " ").strip()
+        srt_entries.append(f"{idx}\n{start_ts} --> {end_ts}\n{clean_text}")
+    return "\n\n".join(srt_entries) + "\n" if srt_entries else ""
+
+
+def generate_txt(segments: List[dict]) -> str:
+    """Generate plain text transcript with timestamps from segments."""
+    lines = []
+    for seg in segments:
+        start_sec = float(seg.get("start_seconds", 0))
+        text = seg.get("text", "").strip()
+        if not text:
+            continue
+        ts = seconds_to_srt_timestamp(start_sec)
+        lines.append(f"[{ts}] {text}")
+    return "\n".join(lines) + "\n" if lines else ""
+
+
+def generate_json_export(session: dict, segments: List[dict],
+                         summaries: List[dict],
+                         canvas_events: List[dict]) -> str:
+    """Generate structured JSON export with segments, metadata, and canvas events."""
+    # Build clean segment list (exclude internal DB fields)
+    clean_segments = []
+    for seg in segments:
+        clean_segments.append({
+            "sequence_index": seg.get("sequence_index"),
+            "text": seg.get("text", ""),
+            "start_seconds": float(seg.get("start_seconds", 0)),
+            "end_seconds": float(seg.get("end_seconds", 0)),
+            "is_final": seg.get("is_final", True),
+            "speaker_label": seg.get("speaker_label"),
+            "keyword_matches": seg.get("keyword_matches"),
+        })
+
+    # Build clean summary list
+    clean_summaries = []
+    for s in summaries:
+        clean_summaries.append({
+            "id": s.get("id"),
+            "summary_type": s.get("summary_type"),
+            "content": s.get("content", ""),
+            "llm_provider": s.get("llm_provider"),
+            "llm_model": s.get("llm_model"),
+            "created_at": s.get("created_at"),
+        })
+
+    # Build clean canvas events list
+    clean_events = []
+    for ev in canvas_events:
+        clean_events.append({
+            "id": ev.get("id"),
+            "event_type": ev.get("event_type"),
+            "session_elapsed_seconds": float(ev.get("session_elapsed_seconds", 0)) if ev.get("session_elapsed_seconds") else None,
+            "timestamp": ev.get("timestamp"),
+            "event_payload": ev.get("event_payload", {}),
+        })
+
+    export_data = {
+        "session": {
+            "id": session.get("id"),
+            "title": session.get("title"),
+            "canvas_type": session.get("canvas_type"),
+            "started_at": session.get("started_at"),
+            "ended_at": session.get("ended_at"),
+            "duration_seconds": session.get("duration_seconds"),
+            "timetable_period_id": session.get("timetable_period_id"),
+            "timetable_event_type": session.get("timetable_event_type"),
+            "timetable_event_label": session.get("timetable_event_label"),
+            "auto_tagged": session.get("auto_tagged", False),
+            "llm_provider": session.get("llm_provider"),
+            "llm_model": session.get("llm_model"),
+            "word_count": session.get("word_count", 0),
+            "segment_count": session.get("segment_count", 0),
+        },
+        "segments": clean_segments,
+        "summaries": clean_summaries,
+        "canvas_events": clean_events,
+    }
+
+    return json.dumps(export_data, indent=2, default=str)
+
+
+def sanitize_filename(name: str) -> str:
+    """Remove or replace characters that are unsafe in filenames."""
+    safe = "".join(c if c.isalnum() or c in " _-." else "_" for c in name)
+    return safe[:100] if safe else "export"
+
+
 @router.post("/sessions", response_model=TranscriptionSessionResponse)
 async def create_session(
     session_data: TranscriptionSessionCreate,
@@ -211,28 +330,86 @@ async def generate_summary(
     summary_request: SummaryGenerateRequest,
     user_id: str = Depends(get_user_id),
 ):
-    """Generate a summary for a session (Phase 1 stub)."""
+    """Generate a summary for a session using the specified LLM provider.
+
+    Phase 3: Full implementation — calls the pluggable LLM client with
+    prompt templates from prompts.py. API key is passed per-request and
+    never stored or logged.
+    """
     supabase = get_supabase_client()
     
     # Verify session exists and user owns it
-    session_check = supabase.supabase.table("transcription_sessions").select("id").eq("id", session_id).eq("user_id", user_id).execute()
+    session_check = supabase.supabase.table("transcription_sessions").select("*").eq("id", session_id).eq("user_id", user_id).execute()
     
     if not session_check.data:
         raise HTTPException(status_code=404, detail="Session not found")
     
-    # Phase 1 stub: TODO implement LLM call in Phase 3
-    content = "[TODO: Generate summary via LLM — provider={}, model={}]".format(
-        summary_request.provider, summary_request.model
-    )
+    session = session_check.data[0]
+    
+    # Build transcript from segments (or use segment range)
+    segments_query = supabase.supabase.table("transcription_segments").select("*").eq("session_id", session_id).order("sequence_index")
+    segments_result = segments_query.execute()
+    
+    if not segments_result.data:
+        raise HTTPException(status_code=400, detail="No segments found for this session")
+    
+    # Apply segment range filter if specified
+    segments = segments_result.data
+    if summary_request.segment_range and len(summary_request.segment_range) == 2:
+        start_idx, end_idx = summary_request.segment_range[0], summary_request.segment_range[1]
+        if start_idx is not None and end_idx is not None:
+            segments = segments[start_idx:end_idx]
+        elif start_idx is not None:
+            segments = segments[start_idx:]
+        elif end_idx is not None:
+            segments = segments[:end_idx]
+    
+    # Build full transcript text from segments
+    transcript_parts = [s["text"] for s in segments if s.get("text")]
+    transcript = "\n".join(transcript_parts)
+    
+    if not transcript.strip():
+        raise HTTPException(status_code=400, detail="Transcript is empty — cannot generate summary")
+    
+    # Build prompt from template
+    system_prompt, user_message = build_prompt(summary_request.summary_type, transcript)
+    
+    # Call the LLM client
+    try:
+        llm_result = await call_llm(
+            provider=summary_request.provider,
+            model=summary_request.model,
+            api_key=summary_request.api_key,
+            system_prompt=system_prompt,
+            user_message=user_message,
+        )
+    except Exception as e:
+        logger.error(f"LLM call failed: {e}")
+        raise HTTPException(status_code=502, detail=f"LLM generation failed: {str(e)}")
+    
+    # Determine segment range for storage
+    seg_start = None
+    seg_end = None
+    if summary_request.segment_range and len(summary_request.segment_range) == 2:
+        seg_start = summary_request.segment_range[0]
+        seg_end = summary_request.segment_range[1]
+    
+    # Build the prompt that was used (for audit trail)
+    prompt_used = f"{system_prompt}\n\n{user_message}" if summary_request.summary_type != "segment" else user_message
     
     # Save summary to database
     summary_data = {
         "session_id": session_id,
         "user_id": user_id,
         "summary_type": summary_request.summary_type,
-        "content": content,
+        "content": llm_result.content,
+        "prompt_used": prompt_used,
         "llm_provider": summary_request.provider,
         "llm_model": summary_request.model,
+        "input_tokens": llm_result.input_tokens,
+        "output_tokens": llm_result.output_tokens,
+        "segment_range_start": seg_start,
+        "segment_range_end": seg_end,
     }
     
     result = supabase.supabase.table("transcription_summaries").insert(summary_data).execute()
@@ -268,26 +445,65 @@ async def export_session(
     export_format: ExportFormat,
     user_id: str = Depends(get_user_id),
 ):
-    """Export session as SRT, TXT, or JSON (Phase 1 stub)."""
+    """Export session as SRT, TXT, or JSON file download.
+
+    Phase 3E: Full implementation — generates properly formatted files
+    and returns them as downloadable responses. API keys are never stored
+    or logged during export.
+    """
     supabase = get_supabase_client()
     
     # Verify ownership
-    session_check = supabase.supabase.table("transcription_sessions").select("id").eq("id", session_id).eq("user_id", user_id).execute()
+    session_check = supabase.supabase.table("transcription_sessions").select("*").eq("id", session_id).eq("user_id", user_id).execute()
     
     if not session_check.data:
         raise HTTPException(status_code=404, detail="Session not found")
     
+    session = session_check.data[0]
+    
     # Get segments
     segments_result = supabase.supabase.table("transcription_segments").select("*").eq("session_id", session_id).order("sequence_index").execute()
     segments = segments_result.data
     
-    if export_format.format == "srt":
-        # Phase 1 stub — implement in Phase 3
-        return {"format": "srt", "content": "[TODO: Generate SRT from segments]"}
-    elif export_format.format == "txt":
-        text = "\n".join(s["text"] for s in segments)
-        return {"format": "txt", "content": text}
-    elif export_format.format == "json":
-        return {"format": "json", "content": {"segments": segments}}
+    # Get summaries (for JSON export)
+    summaries_result = supabase.supabase.table("transcription_summaries").select("*").eq("session_id", session_id).execute()
+    summaries = summaries_result.data
+    
+    # Get canvas events (for JSON export)
+    canvas_result = supabase.supabase.table("canvas_events").select("*").eq("session_id", session_id).order("timestamp").execute()
+    canvas_events = canvas_result.data
+    
+    fmt = export_format.format.lower()
+    
+    if fmt == "srt":
+        content = generate_srt(segments)
+        filename = f"{sanitize_filename(session.get('title', session_id))}_{session.get('started_at', 'export')[:10]}.srt"
+        return FileResponse(
+            io.BytesIO(content.encode("utf-8")),
+            media_type="application/x-subrip",
+            filename=filename,
+            headers={"Content-Disposition": f"attachment; filename*=UTF-8''{filename}"},
+        )
+    
+    elif fmt == "txt":
+        content = generate_txt(segments)
+        filename = f"{sanitize_filename(session.get('title', session_id))}_{session.get('started_at', 'export')[:10]}.txt"
+        return FileResponse(
+            io.BytesIO(content.encode("utf-8")),
+            media_type="text/plain",
+            filename=filename,
+            headers={"Content-Disposition": f"attachment; filename*=UTF-8''{filename}"},
+        )
+    
+    elif fmt == "json":
+        content = generate_json_export(session, segments, summaries, canvas_events)
+        filename = f"{sanitize_filename(session.get('title', session_id))}_{session.get('started_at', 'export')[:10]}.json"
+        return FileResponse(
+            io.BytesIO(content.encode("utf-8")),
+            media_type="application/json",
+            filename=filename,
+            headers={"Content-Disposition": f"attachment; filename*=UTF-8''{filename}"},
+        )
+    
     else:
         raise HTTPException(status_code=400, detail=f"Unsupported format: {export_format.format}")