This article is published in English.
Practical notes: Stop Watching YouTube Videos. Build AI Agents to Start
Operable walkthrough of Practical notes: Stop Watching YouTube Videos. Build AI Agents to Start: contracts, checks, and drop-in code slots for teams shipping this pattern.
The following notes reconstruct a practical path around “Stop Watching YouTube Videos. Build AI Agents to Start Chatting With Them.”. Emphasis stays on contracts, checks, and drop-in code placeholders rather than motivational framing. When working through the Overview stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph.
Introduction
The Introduction stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts.
What this project does
The What this project does stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts.
Architecture Overview
The Architecture Overview stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts. The Architecture Overview stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph.
Core Components
For the Core Components stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.
Setup:
For the Setup stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.
python3 -m venv .venv
source .venv/bin/activate
python -m venv .venv
.venv\Scripts\activate
streamlit>=1.35.0
openai>=1.30.0
python-dotenv>=1.0.0
youtube-transcript-api>=0.6.2
faiss-cpu>=1.8.0
numpy>=1.26.0
requests>=2.31.0
rank-bm25>=0.2.2
#terminal
pip install requirements.txt
Step 1: Fetching Transcript and Chunking
For the Step 1 Fetching Transcript stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.
1.1 Extracting the YouTube video ID
For the 1 1 Extracting the stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.
def extract_video_id(url: str) -> str | None:
patterns = [
r"(?:v=|\/)([0-9A-Za-z_-]{11}).*",
r"(?:youtu\.be\/)([0-9A-Za-z_-]{11})",
r"(?:embed\/)([0-9A-Za-z_-]{11})",
r"^([0-9A-Za-z_-]{11})quot;,
]
for pattern in patterns:
match = re.search(pattern, url)
if match:
return match.group(1)
return None
1.2: Fetching the transcript
For the 1 2 Fetching the stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.
fetched = api.fetch(video_id, languages=['en'])
[{"text": …, "start": …, "duration": …}, …]
1.3: Chunking the transcript with timestamps
For the 1 3 Chunking the stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.
def chunk_transcript(transcript: list[dict], chunk_size: int = 300, overlap: int = 50) -> list[dict]:
chunks = []
words_buffer = []
word_timestamps = []
for entry in transcript:
words = entry['text'].split()
start = entry['start']
duration = entry.get('duration', 2.0)
for i, word in enumerate(words):
t = start + (duration * i / max(len(words), 1))
words_buffer.append(word)
word_timestamps.append(t)
step = chunk_size - overlap
i = 0
while i < len(words_buffer):
end_idx = min(i + chunk_size, len(words_buffer))
chunk_words = words_buffer[i:end_idx]
chunk_times = word_timestamps[i:end_idx]
chunks.append({
"chunk_id": chunk_id,
"text": " ".join(chunk_words),
"start_time": chunk_times[0],
"end_time": chunk_times[-1],
})
i += step
"""
transcript.py - Fetch and parse YouTube transcripts with timestamps
Compatible with youtube-transcript-api v1.x
"""
from youtube_transcript_api import YouTubeTranscriptApi
from youtube_transcript_api._errors import (
NoTranscriptFound, TranscriptsDisabled, VideoUnavailable,
CouldNotRetrieveTranscript,
)
import re
def extract_video_id(url: str) -> str | None:
"""Extract video ID from various YouTube URL formats."""
patterns = [
r"(?:v=|\/)([0-9A-Za-z_-]{11}).*",
r"(?:youtu\.be\/)([0-9A-Za-z_-]{11})",
r"(?:embed\/)([0-9A-Za-z_-]{11})",
r"^([0-9A-Za-z_-]{11})quot;,
]
for pattern in patterns:
match = re.search(pattern, url)
if match:
return match.group(1)
return None
def fetch_transcript(video_id: str) -> list[dict]:
"""
Fetch transcript for a YouTube video.
Returns list of {text, start, duration} dicts.
Compatible with youtube-transcript-api v1.x (instance-based API).
"""
api = YouTubeTranscriptApi()
# Try English first
try:
fetched = api.fetch(video_id, languages=['en'])
return [{"text": s.text, "start": s.start, "duration": s.duration} for s in fetched]
except TranscriptsDisabled:
raise ValueError("Transcripts are disabled for this video.")
except VideoUnavailable:
raise ValueError("Video is unavailable or private.")
except (NoTranscriptFound, CouldNotRetrieveTranscript):
pass # Will try other languages below
except Exception:
pass # Will try other languages below
# Fallback: discover all available languages, use the first one
try:
transcript_list = api.list(video_id)
available = [t.language_code for t in transcript_list]
if not available:
raise ValueError("No transcripts available for this video.")
fetched = api.fetch(video_id, languages=available)
return [{"text": s.text, "start": s.start, "duration": s.duration} for s in fetched]
except TranscriptsDisabled:
raise ValueError("Transcripts are disabled for this video.")
except VideoUnavailable:
raise ValueError("Video is unavailable or private.")
except ValueError:
raise
except Exception as e:
raise ValueError(f"Could not fetch transcript: {str(e)}")
def chunk_transcript(transcript: list[dict], chunk_size: int = 300, overlap: int = 50) -> list[dict]:
"""
Chunk transcript into overlapping windows, preserving timestamps.
Each chunk: {text, start_time, end_time, chunk_id}
"""
chunks = []
words_buffer = []
word_timestamps = []
# Flatten transcript into word-level with timestamps
for entry in transcript:
words = entry['text'].split()
start = entry['start']
duration = entry.get('duration', 2.0)
for i, word in enumerate(words):
t = start + (duration * i / max(len(words), 1))
words_buffer.append(word)
word_timestamps.append(t)
# Slide window
step = chunk_size - overlap
chunk_id = 0
i = 0
while i < len(words_buffer):
end_idx = min(i + chunk_size, len(words_buffer))
chunk_words = words_buffer[i:end_idx]
chunk_times = word_timestamps[i:end_idx]
chunk_text = " ".join(chunk_words)
start_time = chunk_times[0]
end_time = chunk_times[-1]
chunks.append({
"chunk_id": chunk_id,
"text": chunk_text,
"start_time": start_time,
"end_time": end_time,
})
chunk_id += 1
i += step
if end_idx == len(words_buffer):
break
return chunks
def format_timestamp(seconds: float) -> str:
"""Convert seconds to MM:SS or HH:MM:SS string."""
seconds = int(seconds)
h = seconds // 3600
m = (seconds % 3600) // 60
s = seconds % 60
if h > 0:
return f"{h}:{m:02d}:{s:02d}"
return f"{m}:{s:02d}"
def make_youtube_link(video_id: str, seconds: float) -> str:
"""Create a deep-link YouTube URL at a specific timestamp."""
t = int(seconds)
return f"https://www.youtube.com/watch?v={video_id}&t={t}s"
"""
metadata.py - Fetch YouTube video metadata (title, thumbnail, duration, channel)
"""
import requests
import json
import re
def fetch_metadata(video_id: str) -> dict:
"""
Fetch video metadata using YouTube oEmbed API + noembed fallback.
Returns dict with title, author, thumbnail_url, duration_str.
"""
# Try YouTube oEmbed (no API key needed)
oembed_url = f"https://www.youtube.com/oembed?url=https://www.youtube.com/watch?v={video_id}&format=json"
try:
resp = requests.get(oembed_url, timeout=8)
if resp.status_code == 200:
data = resp.json()
thumbnail = f"https://img.youtube.com/vi/{video_id}/mqdefault.jpg"
return {
"title": data.get("title", "Unknown Title"),
"author": data.get("author_name", "Unknown Channel"),
"thumbnail_url": thumbnail,
"video_id": video_id,
"url": f"https://www.youtube.com/watch?v={video_id}",
}
except Exception:
pass
# Fallback: minimal info
return {
"title": f"Video ({video_id})",
"author": "Unknown",
"thumbnail_url": f"https://img.youtube.com/vi/{video_id}/mqdefault.jpg",
"video_id": video_id,
"url": f"https://www.youtube.com/watch?v={video_id}",
}
Step 2: Embedding the Transcript
For the Step 2 Embedding the stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Cite the passages that actually grounded the answer. Without citations, operators cannot tell hallucination from an indexing gap.
2.1: Embedding and FAISS indexing
For the 2 1 Embedding and stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Cite the passages that actually grounded the answer. Without citations, operators cannot tell hallucination from an indexing gap. For the 2 1 Embedding and stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph.
embeddings = get_embeddings(texts, client)
norms = np.linalg.norm(embeddings, axis=1, keepdims=True)
embeddings = embeddings / (norms + 1e-10)
index = faiss.IndexFlatIP(dim)
index.add(embeddings)
def search_index(query, index, chunks, client, top_k=5):
response = client.embeddings.create(model=EMBEDDING_MODEL, input=[query])
q_emb = np.array([response.data[0].embedding], dtype=np.float32)
q_emb = q_emb / (np.linalg.norm(q_emb) + 1e-10)
scores, indices = index.search(q_emb, top_k)
2.2: Building a keyword index with BM25
When working through the 2 2 Building a stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Checkpoint after expensive steps. Resume should not re-bill the same LLM call when an operator retries a later node.
self.tokenized_corpus = [self._tokenize(text) for text in self.corpus]
self.bm25 = BM25Okapi(self.tokenized_corpus)
query_tokens = self._tokenize(query)
scores = self.bm25.get_scores(query_tokens)
"""
embedder.py - Embed transcript chunks and build a FAISS index
"""
import numpy as np
import faiss
import pickle
import os
from openai import OpenAI
EMBEDDING_MODEL = "text-embedding-3-small"
EMBED_BATCH_SIZE = 64
def get_embeddings(texts: list[str], client: OpenAI) -> np.ndarray:
"""Embed a list of texts using OpenAI embeddings in batches."""
all_embeddings = []
for i in range(0, len(texts), EMBED_BATCH_SIZE):
batch = texts[i:i + EMBED_BATCH_SIZE]
response = client.embeddings.create(model=EMBEDDING_MODEL, input=batch)
batch_embeddings = [item.embedding for item in response.data]
all_embeddings.extend(batch_embeddings)
return np.array(all_embeddings, dtype=np.float32)
def build_index(chunks: list[dict], client: OpenAI) -> tuple[faiss.Index, list[dict]]:
"""
Build a FAISS flat L2 index from transcript chunks.
Returns (index, chunks) — chunks are stored as metadata alongside index.
"""
texts = [c["text"] for c in chunks]
embeddings = get_embeddings(texts, client)
# Normalize for cosine similarity
norms = np.linalg.norm(embeddings, axis=1, keepdims=True)
embeddings = embeddings / (norms + 1e-10)
dim = embeddings.shape[1]
index = faiss.IndexFlatIP(dim) # Inner product = cosine after normalization
index.add(embeddings)
return index, chunks
def search_index(
query: str,
index: faiss.Index,
chunks: list[dict],
client: OpenAI,
top_k: int = 5,
) -> list[dict]:
"""
Search the FAISS index for chunks most relevant to query.
Returns top_k chunks with similarity scores.
"""
response = client.embeddings.create(model=EMBEDDING_MODEL, input=[query])
q_emb = np.array([response.data[0].embedding], dtype=np.float32)
# Normalize
q_emb = q_emb / (np.linalg.norm(q_emb) + 1e-10)
scores, indices = index.search(q_emb, top_k)
results = []
for score, idx in zip(scores[0], indices[0]):
if idx < len(chunks):
chunk = chunks[idx].copy()
chunk["score"] = float(score)
results.append(chunk)
return results
"""
keyword_index.py — BM25 keyword indexing and retrieval for transcript chunks.
Provides efficient keyword-based search as complement to semantic vector search.
"""
from rank_bm25 import BM25Okapi
from typing import List, Dict
class KeywordIndex:
"""Build and search a BM25 keyword index from transcript chunks."""
def __init__(self, chunks: List[Dict]):
"""
Initialize BM25 index from chunks.
Args:
chunks: List of chunk dicts with 'text', 'start_time', 'end_time' keys
"""
self.chunks = chunks
self.corpus = [chunk["text"] for chunk in chunks]
# Tokenize: split on whitespace, lowercase, simple punctuation removal
self.tokenized_corpus = [self._tokenize(text) for text in self.corpus]
self.bm25 = BM25Okapi(self.tokenized_corpus)
@staticmethod
def _tokenize(text: str) -> List[str]:
"""Simple tokenization: lowercase, split on whitespace."""
import re
# Convert to lowercase, split on whitespace, remove punctuation
tokens = re.findall(r'\w+', text.lower())
return tokens
def search(self, query: str, top_k: int = 5) -> List[Dict]:
"""
Search BM25 index for chunks matching query.
Args:
query: Search query string
top_k: Number of top results to return
Returns:
List of chunks with 'bm25_score' added; sorted by score descending
"""
query_tokens = self._tokenize(query)
# BM25 returns scores for each document in corpus
scores = self.bm25.get_scores(query_tokens)
# Sort by score descending, get top_k indices
top_indices = sorted(
range(len(scores)),
key=lambda i: scores[i],
reverse=True
)[:top_k]
results = []
for idx in top_indices:
if idx < len(self.chunks):
chunk = self.chunks[idx].copy()
chunk["bm25_score"] = float(scores[idx])
results.append(chunk)
return results
def build_keyword_index(chunks: List[Dict]) -> KeywordIndex:
"""
Convenience function to build a BM25 index from chunks.
Args:
chunks: List of chunk dicts
Returns:
KeywordIndex instance ready for search
"""
return KeywordIndex(chunks)
Step 3: Hybrid retrieval with RRF fusion
When working through the Step 3 Hybrid retrieval stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Measure recall on a fixed question set before tuning prompts. Prompt churn rarely fixes a weak retrieval surface.
score = sum(1 / (rank + k))
Step 4: Query routing — global vs specific
When working through the Step 4 Query routing stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Checkpoint after expensive steps. Resume should not re-bill the same LLM call when an operator retries a later node. When working through the Step 4 Query routing stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph.
response = client.chat.completions.create(
model=MODEL,
messages=[
{"role": "system", "content": ROUTER_PROMPT},
{"role": "user", "content": user_message},
],
)
"""
retrieval_fusion.py — Reciprocal Rank Fusion (RRF) for hybrid search.
Merges keyword (BM25) and semantic (vector) results into a single ranked list.
Uses reciprocal rank fusion formula: score = sum(1 / (rank + k))
"""
from typing import List, Dict, Tuple
def reciprocal_rank_fusion(
keyword_results: List[Dict],
vector_results: List[Dict],
k: int = 60,
) -> List[Dict]:
"""
Merge and rank results from keyword and vector searches using RRF.
RRF formula for each result:
score = sum(1 / (rank_keyword + k) + 1 / (rank_vector + k))
Where rank is 0-indexed position in each result list.
Results appearing in both lists get scores from both; results in one list only
contribute their single score.
Args:
keyword_results: List of chunks from BM25 search (with 'bm25_score' field)
vector_results: List of chunks from vector search (with 'score' field for cosine similarity)
k: RRF parameter; higher k diminishes effect of rank position
Default 60 is standard; tune based on result quality
Returns:
List of unique chunks sorted by fused RRF score (descending)
Each chunk has 'rrf_score', 'keyword_rank', 'vector_rank' fields added
"""
# Build ranking maps: chunk_id -> (rank, original_chunk_dict)
# Using (start_time, end_time) as unique chunk ID
keyword_ranks: Dict[Tuple, Tuple[int, Dict]] = {}
vector_ranks: Dict[Tuple, Tuple[int, Dict]] = {}
for rank, chunk in enumerate(keyword_results):
chunk_id = (chunk.get("start_time"), chunk.get("end_time"))
keyword_ranks[chunk_id] = (rank, chunk)
for rank, chunk in enumerate(vector_results):
chunk_id = (chunk.get("start_time"), chunk.get("end_time"))
vector_ranks[chunk_id] = (rank, chunk)
# Compute RRF scores for all unique chunks
rrf_scores: Dict[Tuple, float] = {}
all_chunks: Dict[Tuple, Dict] = {}
chunk_metadata: Dict[Tuple, Dict] = {} # Track rank info
# Process keyword results
for chunk_id, (rank, chunk) in keyword_ranks.items():
rrf_scores[chunk_id] = 1.0 / (rank + k)
all_chunks[chunk_id] = chunk
chunk_metadata[chunk_id] = {"keyword_rank": rank, "vector_rank": None}
# Process vector results
for chunk_id, (rank, chunk) in vector_ranks.items():
vector_contribution = 1.0 / (rank + k)
if chunk_id in rrf_scores:
rrf_scores[chunk_id] += vector_contribution
chunk_metadata[chunk_id]["vector_rank"] = rank
else:
rrf_scores[chunk_id] = vector_contribution
all_chunks[chunk_id] = chunk
chunk_metadata[chunk_id] = {"keyword_rank": None, "vector_rank": rank}
# Sort by RRF score descending
sorted_chunks = sorted(
all_chunks.items(),
key=lambda item: rrf_scores[item[0]],
reverse=True
)
# Build result list with metadata
results = []
for chunk_id, chunk in sorted_chunks:
result_chunk = chunk.copy()
result_chunk["rrf_score"] = rrf_scores[chunk_id]
result_chunk["keyword_rank"] = chunk_metadata[chunk_id]["keyword_rank"]
result_chunk["vector_rank"] = chunk_metadata[chunk_id]["vector_rank"]
results.append(result_chunk)
return results
def fuse_and_get_top_k(
keyword_results: List[Dict],
vector_results: List[Dict],
top_k: int = 5,
rrf_k: int = 60,
) -> List[Dict]:
"""
Convenience function: fuse results and return top_k.
Args:
keyword_results: BM25 search results
vector_results: Vector search results
top_k: Number of results to return from fused list
rrf_k: RRF parameter
Returns:
Top k chunks from fused ranking
"""
fused = reciprocal_rank_fusion(keyword_results, vector_results, k=rrf_k)
return fused[:top_k]
Step 6: Constructing the prompt
The Step 6 Constructing the stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices.
[MM:SS](https://www.youtube.com/watch?v={video_id}&t=Xs)
"""
chat.py — LLM-routed chat with inline timestamp citations.
Router: one fast GPT call → "global" | "rag"
Global: full transcript passed as context (up to 80k tokens)
RAG: hybrid retrieval (BM25 keyword + vector semantic) with RRF fusion,
LLM cites 1-2 timestamps inline in answer
"""
from openai import OpenAI
from embedder import search_index
from keyword_index import build_keyword_index
from retrieval_fusion import fuse_and_get_top_k
from transcript import format_timestamp, make_youtube_link
import faiss
import json
MODEL = "gpt-4o-mini"
MAX_FULL_TRANSCRIPT_WORDS = 60_000
# ── System prompts ─────────────────────────────────────────────────────────────
ROUTER_PROMPT = """You are a query classifier for a YouTube video Q&A assistant.
Classify the user's question as one of two types:
"global" — The question requires understanding the ENTIRE video.
Examples: summarize, overview, main topics, key takeaways,
chapters, structure, what is this video about, full recap.
"rag" — The question is about a SPECIFIC fact, moment, person, concept,
or timestamp in the video. Examples: when did X happen,
what did the speaker say about Y, explain concept Z.
Reply with ONLY a JSON object: {"route": "global"} or {"route": "rag"}
No explanation. No other text."""
GLOBAL_SYSTEM_PROMPT = """You are an intelligent video assistant. You have been given the COMPLETE transcript of a YouTube video with timestamps.
Instructions:
- Answer the user's question using the full transcript comprehensively.
- For summaries: cover ALL major sections, not just the beginning.
- For "main sections/topics": identify distinct topic shifts and list each with its start timestamp.
- Cite timestamps inline using this EXACT markdown format: [MM:SS](https://www.youtube.com/watch?v={video_id}&t=Xs)
where X is the timestamp in seconds. Always include the 's' suffix after the number. Example: &t=315s not &t=315
- Be well-structured — use numbered lists or clear sections.
- Only cite timestamps that are genuinely relevant to that point.
"""
SPECIFIC_SYSTEM_PROMPT = """You are an intelligent video assistant. You have been given relevant excerpts from a YouTube video transcript.
Instructions:
- Answer the question based ONLY on the provided transcript excerpts.
- Cite 1 to 3 timestamps INLINE in your answer using this EXACT markdown format:
[MM:SS](https://www.youtube.com/watch?v={video_id}&t=Xs)
where X is the timestamp in seconds. Always include the 's' suffix. Example: &t=315s not &t=315
- Only cite a timestamp when it directly supports the specific sentence you are writing.
- Do NOT list all timestamps at the end — weave them naturally into the answer.
- If the answer spans multiple parts of the video, show each part as a numbered point with its own inline timestamp.
- If the context does not contain the answer, say: "I couldn't find information about that in this video."
- Never make up information not present in the provided context.
"""
# ── LLM Router ────────────────────────────────────────────────────────────────
def classify_query(user_message: str, client: OpenAI) -> str:
"""
Ask GPT-4o-mini to classify the query as 'global' or 'rag'.
Falls back to 'rag' on any error.
"""
try:
response = client.chat.completions.create(
model=MODEL,
messages=[
{"role": "system", "content": ROUTER_PROMPT},
{"role": "user", "content": user_message},
],
temperature=0,
max_tokens=20,
)
raw = response.choices[0].message.content.strip()
parsed = json.loads(raw)
route = parsed.get("route", "rag")
return route if route in ("global", "rag") else "rag"
except Exception:
return "rag" # safe default
# ── Context builders ──────────────────────────────────────────────────────────
def build_full_transcript_context(all_chunks: list[dict]) -> str:
"""Concatenate ALL chunks sorted by time, capped at MAX_FULL_TRANSCRIPT_WORDS."""
sorted_chunks = sorted(all_chunks, key=lambda c: c["start_time"])
parts = []
words_so_far = 0
for chunk in sorted_chunks:
chunk_words = len(chunk["text"].split())
if words_so_far + chunk_words > MAX_FULL_TRANSCRIPT_WORDS:
parts.append("[... transcript truncated for length ...]")
break
ts = format_timestamp(chunk["start_time"])
parts.append(f"[{ts}] {chunk['text']}")
words_so_far += chunk_words
return "\n".join(parts)
def build_rag_context(chunks: list[dict]) -> str:
"""Format retrieved chunks into a timestamped context block."""
parts = []
for chunk in chunks:
ts = format_timestamp(chunk["start_time"])
end_ts = format_timestamp(chunk["end_time"])
parts.append(f"[{ts} - {end_ts}]\n{chunk['text']}")
return "\n\n---\n\n".join(parts)
# ── Source extractor (parses inline links from LLM reply) ────────────────────
def extract_sources_from_reply(reply: str, video_id: str) -> list[dict]:
"""
Parse timestamp markdown links that the LLM wrote inline.
Matches patterns like [4:32](https://...&t=272s)
Returns deduplicated list of {timestamp, seconds, link}.
"""
import re
# Match [MM:SS] or [H:MM:SS] followed by a YouTube URL with &t=Xs
# s suffix is optional — LLM sometimes writes &t=315 not &t=315s
pattern = r'\[([\d]{1,2}:\d{2}(?::\d{2})?)\]\((https://www\.youtube\.com/watch\?v=[\w-]+&t=(\d+)s?)\)'
matches = re.findall(pattern, reply)
seen = set()
sources = []
for ts_label, url, seconds_str in matches:
seconds = int(seconds_str)
if seconds not in seen:
seen.add(seconds)
sources.append({
"timestamp": ts_label,
"seconds": float(seconds),
"link": url,
})
return sources
# ── Main chat function ────────────────────────────────────────────────────────
def chat_with_video(
user_message: str,
conversation_history: list[dict],
index: faiss.Index,
chunks: list[dict],
video_id: str,
client: OpenAI,
top_k: int = 5,
keyword_index=None,
) -> tuple[str, list[dict]]:
"""
1. Classify query → global | rag
2. Build context accordingly
3. For 'rag': use hybrid retrieval (BM25 + semantic with RRF fusion)
4. Call GPT-4o-mini with inline-timestamp instructions
5. Parse timestamps from reply for UI chips
Args:
user_message: User's query
conversation_history: Previous messages in conversation
index: FAISS vector index
chunks: All transcript chunks
video_id: YouTube video ID
client: OpenAI client
top_k: Number of results to return after fusion
keyword_index: KeywordIndex instance for BM25 search (optional)
Returns:
Tuple of (reply_text, sources)
"""
# ── Step 1: Route ──────────────────────────────────────────────────────
route = classify_query(user_message, client)
# ── Step 2: Build context ──────────────────────────────────────────────
if route == "global":
context = build_full_transcript_context(chunks)
system_prompt = GLOBAL_SYSTEM_PROMPT.replace("{video_id}", video_id)
context_label = "FULL VIDEO TRANSCRIPT (with timestamps):"
max_tokens = 1800
else:
# ── Hybrid retrieval: BM25 keyword + semantic vector with RRF fusion ──
# Retrieve top 10 from each method, then fuse to top_k
vector_results = search_index(user_message, index, chunks, client, top_k=10)
if keyword_index is not None:
keyword_results = keyword_index.search(user_message, top_k=10)
# Fuse using Reciprocal Rank Fusion
retrieved = fuse_and_get_top_k(
keyword_results,
vector_results,
top_k=top_k,
rrf_k=60
)
else:
# Fallback: use vector search only if keyword index not available
retrieved = vector_results[:top_k]
context = build_rag_context(retrieved)
system_prompt = SPECIFIC_SYSTEM_PROMPT.replace("{video_id}", video_id)
context_label = "RELEVANT TRANSCRIPT EXCERPTS (hybrid keyword + semantic search):"
max_tokens = 900
# ── Step 3: Call LLM ───────────────────────────────────────────────────
messages = [
{"role": "system", "content": system_prompt},
{"role": "system", "content": f"{context_label}\n\n{context}"},
]
messages.extend(conversation_history)
messages.append({"role": "user", "content": user_message})
response = client.chat.completions.create(
model=MODEL,
messages=messages,
temperature=0,
max_tokens=max_tokens,
)
reply = response.choices[0].message.content
# ── Step 4: Extract inline timestamp links as UI chips ─────────────────
sources = extract_sources_from_reply(reply, video_id)
return reply, sources
Step 7: Extracting sourced timestamps
The Step 7 Extracting sourced stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts.
pattern = r'\[([\d]{1,2}:\d{2}(?::\d{2})?)\]\((https://www\.youtube\.com/watch\?v=[\w-]+&t=(\d+)s?)\)'
Step 8: Streamlit UI and session state
The Step 8 Streamlit UI stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts. The Step 8 Streamlit UI stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph.
@import url('https://fonts.googleapis.com/css2?family=Google+Sans:wght@400;500;600&family=Roboto:wght@300;400;500&display=swap');
/* Reset & base */
html, body, [class*="css"] {
font-family: 'Roboto', sans-serif;
margin: 0; padding: 0;
}
.stApp {
background: #0f0f0f;
color: #f1f1f1;
}
/* Hide streamlit chrome */
#MainMenu, footer, header { visibility: hidden; }
[data-testid="stSidebar"] { display: none; }
.block-container {
padding: 0 !important;
max-width: 100% !important;
}
/* ── TOP NAV BAR ── */
.topbar {
display: flex;
align-items: center;
justify-content: space-between;
background: #0f0f0f;
border-bottom: 1px solid #272727;
padding: 10px 20px;
position: sticky;
top: 0;
z-index: 100;
}
.topbar-left {
display: flex;
align-items: center;
gap: 12px;
}
.yt-logo {
font-size: 1.3rem;
font-weight: 700;
color: #fff;
letter-spacing: -0.5px;
}
.yt-logo span { color: #ff0000; }
.url-input-wrap {
flex: 1;
max-width: 600px;
margin: 0 24px;
}
/* ── MAIN TWO-PANEL LAYOUT ── */
.main-panels {
display: flex;
height: calc(100vh - 57px);
overflow: hidden;
}
/* Left: video panel */
.video-panel {
flex: 1;
background: #000;
display: flex;
flex-direction: column;
overflow: hidden;
}
.video-embed-wrap {
position: relative;
width: 100%;
padding-top: 56.25%; /* 16:9 */
background: #000;
flex-shrink: 0;
}
.video-embed-wrap iframe {
position: absolute;
top: 0; left: 0;
width: 100%; height: 100%;
border: none;
}
.video-info {
padding: 16px 20px;
border-top: 1px solid #272727;
background: #0f0f0f;
flex-shrink: 0;
}
.video-title {
font-family: 'Roboto', sans-serif;
font-size: 1.1rem;
font-weight: 500;
color: #f1f1f1;
margin: 0 0 4px 0;
line-height: 1.4;
}
.video-channel {
font-size: 0.82rem;
color: #aaa;
margin: 0;
}
/* Right: chat panel */
.chat-panel {
width: 400px;
min-width: 340px;
max-width: 420px;
background: #212121;
border-left: 1px solid #272727;
display: flex;
flex-direction: column;
overflow: hidden;
height: 100%;
}
.chat-header {
padding: 14px 18px 12px;
border-bottom: 1px solid #333;
flex-shrink: 0;
background: #212121;
}
.chat-header-top {
display: flex;
align-items: center;
justify-content: space-between;
margin-bottom: 2px;
}
.chat-title {
font-family: 'Roboto', sans-serif;
font-size: 1rem;
font-weight: 500;
color: #f1f1f1;
margin: 0;
}
.gemini-star {
font-size: 1.1rem;
margin-right: 6px;
}
.chat-subtitle {
font-size: 0.75rem;
color: #aaa;
margin-top: 2px;
}
/* Suggested questions */
.suggestions {
padding: 14px 16px 8px;
border-bottom: 1px solid #2d2d2d;
flex-shrink: 0;
}
.suggestions-label {
font-size: 0.78rem;
color: #aaa;
margin-bottom: 8px;
}
.suggestion-chips {
display: flex;
flex-direction: column;
gap: 6px;
}
.suggestion-chip {
background: transparent;
border: 1px solid #3d3d3d;
border-radius: 18px;
padding: 7px 14px;
font-size: 0.8rem;
color: #c8c8c8;
cursor: pointer;
text-align: right;
width: fit-content;
align-self: flex-end;
transition: background 0.15s, border-color 0.15s;
line-height: 1.3;
}
.suggestion-chip:hover {
background: #2d2d2d;
border-color: #555;
color: #f1f1f1;
}
/* Chat messages area */
.chat-messages {
flex: 1;
overflow-y: auto;
padding: 16px;
display: flex;
flex-direction: column;
gap: 14px;
scrollbar-width: thin;
scrollbar-color: #3d3d3d #212121;
}
/* Thinking dots */
@keyframes thinking-pulse {
0%, 80%, 100% { opacity: 0.2; transform: scale(0.8); }
40% { opacity: 1; transform: scale(1.1); }
}
.thinking-dot {
display: inline-block;
width: 7px; height: 7px;
border-radius: 50%;
background: #666;
animation: thinking-pulse 1.2s ease-in-out infinite;
}
/* Timestamp dropdown chip — pure CSS, no JS */
.ts-dropdown {
position: relative;
display: inline-block;
vertical-align: middle;
margin: 0 2px;
}
.ts-chip {
display: inline-flex;
align-items: center;
gap: 3px;
background: linear-gradient(180deg, #1c2a3a 0%, #142233 100%);
border: 1px solid #2a3f5a;
border-radius: 5px;
padding: 1px 8px;
font-size: 0.78rem;
font-family: 'Roboto Mono', monospace;
color: #8ab4f8;
cursor: pointer;
user-select: none;
white-space: nowrap;
transition: background 0.12s;
box-shadow: 0 1px 4px rgba(0,0,0,0.35);
}
.ts-chip:hover { background: #253549; color: #b0ccff; }
.ts-menu {
display: none;
position: absolute;
bottom: 100%;
margin-bottom: 6px;
left: 0;
background: #141a24;
border: 1px solid #2b3a52;
border-radius: 8px;
min-width: 210px;
z-index: 9999;
overflow: hidden;
box-shadow: 0 10px 24px rgba(0,0,0,0.55);
}
.ts-dropdown:hover .ts-menu,
.ts-dropdown:focus-within .ts-menu {
display: block;
}
.ts-menu::after {
content: "";
position: absolute;
bottom: -6px;
left: 0;
width: 100%;
height: 6px;
}
.ts-option {
display: block;
padding: 9px 14px;
font-size: 0.82rem;
color: #cfd8ea;
text-decoration: none;
cursor: pointer;
transition: background 0.12s;
white-space: nowrap;
}
.ts-option:hover { background: #223248; color: #fff; }
.ts-option + .ts-option { border-top: 1px solid #2d3a50; }
.chat-messages::-webkit-scrollbar { width: 4px; }
.chat-messages::-webkit-scrollbar-track { background: #212121; }
.chat-messages::-webkit-scrollbar-thumb { background: #3d3d3d; border-radius: 2px; }
/* Message bubbles */
.msg-user {
align-self: flex-end;
background: #2d2d2d;
border-radius: 18px 18px 4px 18px;
padding: 10px 14px;
max-width: 85%;
font-size: 0.87rem;
color: #f1f1f1;
line-height: 1.5;
word-wrap: break-word;
}
.msg-ai-wrap {
align-self: flex-start;
max-width: 95%;
display: flex;
flex-direction: column;
gap: 6px;
}
.msg-ai-label {
font-size: 0.72rem;
color: #888;
display: flex;
align-items: center;
gap: 4px;
margin-bottom: 2px;
}
.msg-ai {
background: transparent;
font-size: 0.87rem;
color: #e0e0e0;
line-height: 1.6;
word-wrap: break-word;
}
.msg-ai a {
color: #8ab4f8;
text-decoration: none;
}
.msg-ai a:hover { text-decoration: underline; }
/* Timestamp source chips */
.source-chips {
display: flex;
flex-wrap: wrap;
gap: 5px;
margin-top: 4px;
}
.inline-ts {
display: inline-flex;
align-items: center;
gap: 3px;
background: #1e2a3a;
border: 1px solid #2a3f5a;
border-radius: 5px;
padding: 1px 7px;
font-size: 0.78rem;
font-family: 'Roboto Mono', monospace;
color: #8ab4f8;
cursor: pointer;
transition: background 0.12s, transform 0.1s;
user-select: none;
white-space: nowrap;
vertical-align: middle;
margin: 0 2px;
}
.inline-ts:hover { background: #253549; color: #b0ccff; transform: translateY(-1px); }
.inline-ts:active { transform: translateY(0); background: #2a3f5a; }
/* Chat input area */
.chat-input-area {
padding: 10px 14px 8px;
border-top: 1px solid #2d2d2d;
background: #212121;
flex-shrink: 0;
}
.chat-disclaimer {
text-align: center;
font-size: 0.67rem;
color: #666;
padding: 4px 0 0;
}
/* Streamlit input overrides */
.stTextInput > div > div > input {
background: #2d2d2d !important;
border: 1px solid #3d3d3d !important;
border-radius: 22px !important;
color: #f1f1f1 !important;
font-size: 0.87rem !important;
padding: 10px 18px !important;
font-family: 'Roboto', sans-serif !important;
}
.stTextInput > div > div > input:focus {
border-color: #555 !important;
box-shadow: none !important;
outline: none !important;
}
.stTextInput > div > div > input::placeholder { color: #888 !important; }
.stButton > button {
background: transparent;
border: none;
color: #8ab4f8;
font-size: 0.85rem;
font-weight: 500;
padding: 6px 12px;
border-radius: 4px;
cursor: pointer;
font-family: 'Roboto', sans-serif;
transition: background 0.15s;
}
.stButton > button:hover { background: #2d2d2d; color: #c0d4ff; }
/* Empty / loading states */
.empty-state {
flex: 1;
display: flex;
flex-direction: column;
align-items: center;
justify-content: center;
text-align: center;
padding: 32px 24px;
color: #888;
}
.empty-icon { font-size: 2rem; margin-bottom: 10px; }
.empty-text { font-size: 0.85rem; line-height: 1.6; }
/* Top URL bar inputs */
div[data-testid="stHorizontalBlock"] .stTextInput > div > div > input {
background: #121212 !important;
border: 1px solid #303030 !important;
border-radius: 22px !important;
color: #f1f1f1 !important;
font-size: 0.88rem !important;
padding: 9px 16px !important;
}
/* Spinner */
.stSpinner > div { border-top-color: #aaa !important; }
/* Chips (suggestion buttons) styled via st.button with key trick */
div[data-suggestion="true"] .stButton > button {
background: transparent !important;
border: 1px solid #3d3d3d !important;
border-radius: 18px !important;
color: #c8c8c8 !important;
font-size: 0.8rem !important;
padding: 7px 14px !important;
width: 100% !important;
text-align: right !important;
justify-content: flex-end !important;
}
"""
app.py - YouTube AI Chat — Streamlit frontend
Two-panel layout: embedded video left, chat right
"""
import streamlit as st
import streamlit.components.v1 as components
from openai import OpenAI
import time
import os
from dotenv import load_dotenv
load_dotenv()
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY", "")
from transcript import extract_video_id, fetch_transcript, chunk_transcript, format_timestamp, make_youtube_link
from metadata import fetch_metadata
from embedder import build_index
from keyword_index import build_keyword_index
from chat import chat_with_video
# ── Page config ────────────────────────────────────────────────────────────────
st.set_page_config(
page_title="YT Chat",
page_icon="🎬",
layout="wide",
initial_sidebar_state="collapsed",
)
# ── Custom CSS ─────────────────────────────────────────────────────────────────
def load_local_css(filename: str):
css_path = os.path.join(os.path.dirname(__file__), filename)
if os.path.exists(css_path):
with open(css_path, "r", encoding="utf-8") as f:
st.markdown(f"<style>{f.read()}</style>", unsafe_allow_html=True)
else:
st.warning(f"Missing CSS file: {css_path}")
load_local_css("style.css")
# ── Query param handler is already set up above ──────────────────────────────────────
# ── Session state ──────────────────────────────────────────────────────────────
def init_state():
defaults = {
"videos": {},
"active_video_id": None,
"conversations": {},
"client": None,
"pending_input": "",
"awaiting_answer": "", # question waiting for LLM response
"processing_answer": False,
}
for k, v in defaults.items():
if k not in st.session_state:
st.session_state[k] = v
init_state()
# Handle jump-to-timestamp from HTML button click
if "jump_to" in st.query_params:
try:
jump_seconds = st.query_params.get("jump_to")
if jump_seconds:
st.session_state.jump_to_seconds = int(jump_seconds)
# Clear the param to avoid re-triggering
params = dict(st.query_params)
del params["jump_to"]
st.query_params.clear()
for k, v in params.items():
st.query_params[k] = v
except Exception as e:
print(f"Error handling jump_to param: {e}")
def get_client():
return st.session_state.client
def active_video():
vid = st.session_state.active_video_id
if vid and vid in st.session_state.videos:
return st.session_state.videos[vid]
return None
def active_conversation():
vid = st.session_state.active_video_id
if vid and vid not in st.session_state.conversations:
st.session_state.conversations[vid] = []
if vid:
return st.session_state.conversations[vid]
return []
# ── TOP NAV ────────────────────────────────────────────────────────────────────
st.markdown("""
<div class="topbar">
<div style="font-size:1.25rem;font-weight:700;color:#fff;letter-spacing:-0.3px;">
<span style="color:#ff0000;">▶</span> YT Chat
</div>
</div>
""", unsafe_allow_html=True)
# Controls row below topbar
ctrl_col1, ctrl_col2, ctrl_col3 = st.columns([4, 1, 1])
with ctrl_col1:
yt_url = st.text_input(
"url", placeholder="Paste YouTube URL...",
label_visibility="collapsed", key="url_input"
)
with ctrl_col2:
load_btn = st.button("Load Video", use_container_width=True)
with ctrl_col3:
# Show loaded videos selector if multiple
if len(st.session_state.videos) > 1:
video_options = {v["meta"]["title"][:28] + "…": k
for k, v in st.session_state.videos.items()}
selected_label = st.selectbox(
"Switch", list(video_options.keys()),
label_visibility="collapsed"
)
st.session_state.active_video_id = video_options[selected_label]
elif len(st.session_state.videos) == 1:
st.markdown(
f'<div style="font-size:0.75rem;color:#888;padding:8px 0;">1 video loaded</div>',
unsafe_allow_html=True
)
# Handle load
if load_btn:
url_val = yt_url.strip()
if not OPENAI_API_KEY:
st.error("OPENAI_API_KEY not found in .env file.")
st.stop()
elif not url_val:
st.error("Paste a YouTube URL.")
else:
if not st.session_state.client:
st.session_state.client = OpenAI(api_key=OPENAI_API_KEY)
video_id = extract_video_id(url_val)
if not video_id:
st.error("Couldn't parse a video ID from that URL.")
elif video_id in st.session_state.videos:
st.session_state.active_video_id = video_id
st.success("Already loaded — switched to it.")
st.rerun()
else:
prog = st.progress(0, text="Fetching metadata...")
try:
meta = fetch_metadata(video_id)
prog.progress(15, text="Fetching transcript...")
raw = fetch_transcript(video_id)
prog.progress(40, text="Chunking transcript...")
chunks = chunk_transcript(raw)
prog.progress(60, text=f"Embedding {len(chunks)} chunks...")
index, chunks = build_index(chunks, get_client())
prog.progress(80, text="Building keyword index...")
keyword_index = build_keyword_index(chunks)
prog.progress(95, text="Almost done...")
st.session_state.videos[video_id] = {
"meta": meta, "chunks": chunks,
"index": index, "keyword_index": keyword_index,
"chunk_count": len(chunks),
}
st.session_state.active_video_id = video_id
st.session_state.conversations[video_id] = []
prog.progress(100, text="Ready!")
time.sleep(0.3)
prog.empty()
st.rerun()
except ValueError as e:
prog.empty()
st.error(str(e))
except Exception as e:
prog.empty()
st.error(f"Error: {e}")
st.markdown("<div style='height:1px;background:#272727;margin:0;'></div>", unsafe_allow_html=True)
# ── MAIN TWO-PANEL LAYOUT ──────────────────────────────────────────────────────
video = active_video()
if not video:
# Empty state
st.markdown("""
<div style="display:flex;align-items:center;justify-content:center;
height:calc(100vh - 120px);flex-direction:column;
text-align:center;color:#555;gap:12px;">
<div style="font-size:3rem;">▶</div>
<div style="font-size:1rem;color:#888;font-weight:500;">Paste a YouTube URL above to get started</div>
<div style="font-size:0.82rem;color:#555;max-width:380px;line-height:1.6;">
Chat with any video — answers grounded in the transcript with clickable timestamps
</div>
</div>
""", unsafe_allow_html=True)
else:
meta = video["meta"]
chunks = video["chunks"]
index = video["index"]
video_id = meta["video_id"]
conversation = active_conversation()
# ── Two columns: video | chat ──────────────────────────────────────────────
left_col, right_col = st.columns([1.15, 0.85], gap="small")
# ── LEFT: Video embed + info ───────────────────────────────────────────────
with left_col:
origin = "http://localhost:8501"
# Create placeholder for video panel to allow re-rendering on jump
video_placeholder = st.empty()
# Check if we need to update start time
start_time = 0
if hasattr(st.session_state, 'jump_to_seconds') and st.session_state.jump_to_seconds:
start_time = st.session_state.jump_to_seconds
st.session_state.jump_to_seconds = None # Reset for next jump
with video_placeholder.container():
st.markdown(f"""
<div class="video-panel">
<div class="video-embed-wrap">
<iframe
id="yt-player"
name="yt-player"
src="https://www.youtube.com/embed/{video_id}?rel=0&modestbranding=1&enablejsapi=1&autoplay=1&start={start_time}&origin={origin}"
allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share"
allowfullscreen>
</iframe>
</div>
<div class="video-info">
<div class="video-title">{meta['title']}</div>
<div class="video-channel">{meta['author']}</div>
<div class="meta-row" style="display:flex;align-items:center;gap:12px;">
<span class="badge">{video['chunk_count']} chunks indexed</span>
<a class="yt-link" href="{meta['url']}" target="_blank">↗ Open on YouTube</a>
</div>
</div>
</div>
""", unsafe_allow_html=True)
# ── RIGHT: Chat panel ──────────────────────────────────────────────────────
with right_col:
# Clear input field if flag is set from previous submission
if st.session_state.get("should_clear_input", False):
st.session_state.chat_input = ""
st.session_state.should_clear_input = False
# Chat header
st.markdown("""
<div style="background:#212121;border:1px solid #2d2d2d;border-radius:10px;
overflow:hidden;display:flex;flex-direction:column;">
<div style="padding:14px 18px 10px;border-bottom:1px solid #2d2d2d;">
<div style="display:flex;align-items:center;gap:6px;">
<span style="font-size:1rem;">💬</span>
<span style="font-size:0.95rem;font-weight:500;color:#f1f1f1;">Ask about this video</span>
</div>
<div style="font-size:0.73rem;color:#777;margin-top:2px;">
Answers grounded in transcript · click timestamps to jump
</div>
</div>
</div>
""", unsafe_allow_html=True)
# Suggested questions + chat container
# Height shrinks when suggestions are visible so input stays on screen
suggestions = [ "Summarize this video", "What are the main topics discussed?", "What are the key takeaways?", ]
if st.session_state.get("processing_answer"):
st.markdown(
"<style>div[data-suggestion='true']{display:none !important;}</style>",
unsafe_allow_html=True
)
if not conversation:
is_thinking = bool(st.session_state.get("processing_answer"))
chat_container = st.container(height=230)
with chat_container:
if is_thinking:
question_text = st.session_state.get("awaiting_answer", "")
st.markdown(f"""
<div style="display:flex;justify-content:flex-end;margin:6px 0;">
<div style="background:#2d2d2d;border-radius:18px 18px 4px 18px;
padding:10px 14px;max-width:88%;font-size:0.86rem;
color:#f1f1f1;line-height:1.5;word-wrap:break-word;">
{question_text}
</div>
</div>
<div style="padding:8px 4px;">
<div style="font-size:0.7rem;color:#777;margin-bottom:6px;">✦ AI Assistant</div>
<div style="display:flex;align-items:center;gap:6px;">
<span class="thinking-dot"></span>
<span class="thinking-dot" style="animation-delay:.2s"></span>
<span class="thinking-dot" style="animation-delay:.4s"></span>
<span style="margin-left:4px;font-size:0.8rem;color:#555;">Thinking...</span>
</div>
</div>
""", unsafe_allow_html=True)
else:
st.markdown('<div data-suggestion="true">', unsafe_allow_html=True)
st.markdown("""
<div style="font-size:0.76rem;color:#888;margin-bottom:8px;padding:2px 2px 0;">
Not sure what to ask? Choose something:
</div>
""", unsafe_allow_html=True)
for i, s in enumerate(suggestions):
if st.button(s, key=f"suggestion_{i}", use_container_width=True):
st.session_state.pending_input = s
st.rerun()
st.markdown("""
<div style="text-align:center;color:#444;font-size:0.8rem;
padding:14px 16px 4px;line-height:1.7;">
Hello! Curious about what you're watching?<br>I'm here to help.
</div>
""", unsafe_allow_html=True)
st.markdown("</div>", unsafe_allow_html=True)
else:
# Conversation active — full height container, no suggestions
chat_container = st.container(height=420)
with chat_container:
for turn_idx, turn in enumerate(conversation):
if turn["role"] == "user":
st.markdown(f"""
<div style="display:flex;justify-content:flex-end;margin:6px 0;">
<div style="background:#2d2d2d;border-radius:18px 18px 4px 18px;
padding:10px 14px;max-width:88%;font-size:0.86rem;
color:#f1f1f1;line-height:1.5;word-wrap:break-word;">
{turn['content']}
</div>
</div>
""", unsafe_allow_html=True)
else:
import re as _re
reply_raw = turn["content"]
ts_pattern = r'\[([\d]{1,2}:[\d]{2}(?::[\d]{2})?)\]\((https://www\.youtube\.com/watch\?v=[\w-]+&t=(\d+)s?)\)'
def _timestamp_to_seconds(label: str) -> int:
parts = [int(p) for p in label.split(":")]
if len(parts) == 2:
return parts[0] * 60 + parts[1]
return parts[0] * 3600 + parts[1] * 60 + parts[2]
def _extract_timestamps(text: str):
extracted = []
# 1) Markdown links: [12:34](https://www.youtube.com/watch?v=...&t=754s)
for m in _re.finditer(ts_pattern, text):
extracted.append({"label": m.group(1), "seconds": m.group(3)})
cleaned = _re.sub(ts_pattern, '', text)
# 2) Plain timestamps: (12:34), 12:34, or 1:02:33
plain_ts_pattern = r'(?<!\d)(\d{1,2}:[0-5]\d(?::[0-5]\d)?)(?!\d)'
for m in _re.finditer(plain_ts_pattern, cleaned):
label = m.group(1)
seconds = str(_timestamp_to_seconds(label))
extracted.append({"label": label, "seconds": seconds})
# Remove parenthesized plain timestamps from text once captured.
cleaned = _re.sub(r'\(\s*\d{1,2}:[0-5]\d(?::[0-5]\d)?\s*\)', '', cleaned)
# Deduplicate while preserving order.
deduped = []
seen = set()
for item in extracted:
key = (item["label"], item["seconds"])
if key in seen:
continue
seen.add(key)
deduped.append(item)
return cleaned.strip(), deduped
def _format_block(text: str) -> str:
block = _re.sub(r'\*\*(.+?)\*\*', r'<strong>\1</strong>', text)
block = _re.sub(r'(?m)^\s*(\d+)[.)]\s+', lambda mm: f'<strong>{mm.group(1)}.</strong> ', block)
block = _re.sub(r'(?m)^\s*[-•]\s+', '• ', block)
return block.replace('\n', '<br>')
paragraphs = [p for p in reply_raw.split('\n\n') if p.strip()]
parsed_sections = []
for para in paragraphs:
numbered_starts = list(_re.finditer(r'(?m)^\s*\d+[.)]\s+', para))
# If paragraph contains multiple numbered items, split it into
# per-item sections so each item can get its own timestamp row.
if len(numbered_starts) >= 2:
intro_text = para[:numbered_starts[0].start()].strip()
if intro_text:
intro_display, intro_timestamps = _extract_timestamps(intro_text)
parsed_sections.append({
"text": intro_display,
"timestamps": intro_timestamps,
"is_timestamp_only": bool(intro_timestamps) and not intro_display,
})
for i, match in enumerate(numbered_starts):
start = match.start()
end = numbered_starts[i + 1].start() if i + 1 < len(numbered_starts) else len(para)
item_text = para[start:end].strip()
item_display, item_timestamps = _extract_timestamps(item_text)
parsed_sections.append({
"text": item_display,
"timestamps": item_timestamps,
"is_timestamp_only": bool(item_timestamps) and not item_display,
})
continue
para_display, para_timestamps = _extract_timestamps(para)
parsed_sections.append({
"text": para_display,
"timestamps": para_timestamps,
"is_timestamp_only": bool(para_timestamps) and not para_display,
})
# Redistribute timestamp-only paragraphs when possible so timestamps
# appear under each answer section (especially for numbered lists).
render_sections = []
sec_idx = 0
while sec_idx < len(parsed_sections):
section = parsed_sections[sec_idx]
if section["is_timestamp_only"]:
if render_sections:
render_sections[-1]["timestamps"].extend(section["timestamps"])
sec_idx += 1
continue
lines = [ln.strip() for ln in section["text"].splitlines() if ln.strip()]
numbered_lines = [ln for ln in lines if _re.match(r'^\d+[.)]\s+.+', ln)]
next_is_ts_only = (
sec_idx + 1 < len(parsed_sections)
and parsed_sections[sec_idx + 1]["is_timestamp_only"]
)
if numbered_lines and next_is_ts_only and not section["timestamps"]:
ts_pool = parsed_sections[sec_idx + 1]["timestamps"]
if len(ts_pool) >= len(numbered_lines):
for i, line in enumerate(numbered_lines):
render_sections.append({
"text": line,
"timestamps": [ts_pool[i]],
})
if len(ts_pool) > len(numbered_lines):
render_sections[-1]["timestamps"].extend(ts_pool[len(numbered_lines):])
sec_idx += 2
continue
render_sections.append({
"text": section["text"],
"timestamps": section["timestamps"][:],
})
sec_idx += 1
st.markdown("<div style='margin:6px 0;'><div style='font-size:0.7rem;color:#777;margin-bottom:4px;display:flex;align-items:center;gap:4px;'><span>✦</span> AI Assistant</div></div>", unsafe_allow_html=True)
for para_idx, section in enumerate(render_sections):
if not section["text"] and not section["timestamps"]:
continue
if section["text"]:
para_html = _format_block(section["text"])
st.markdown(
f'<div style="font-size:0.86rem;color:#e0e0e0;line-height:1.6;word-wrap:break-word;margin-bottom:6px;">{para_html}</div>',
unsafe_allow_html=True
)
if section["timestamps"]:
button_cols = st.columns(len(section["timestamps"]), gap='small')
for idx, ts_data in enumerate(section["timestamps"]):
label = ts_data["label"]
seconds = ts_data["seconds"]
with button_cols[idx]:
if st.button(
f'⏱ {label}',
key=f'ts_btn_{video_id}_{turn_idx}_{para_idx}_{idx}_{label}_{seconds}',
use_container_width=True
):
st.session_state.jump_to_seconds = int(seconds)
st.rerun()
# Show thinking bubble inside container if answer is pending
if st.session_state.get("processing_answer"):
st.markdown("""
<div style="margin:6px 0;">
<div style="font-size:0.7rem;color:#777;margin-bottom:4px;
display:flex;align-items:center;gap:4px;">
<span>✦</span> AI Assistant
</div>
<div style="display:flex;align-items:center;gap:6px;
color:#555;font-size:0.82rem;padding:4px 0;">
<span class="thinking-dot"></span>
<span class="thinking-dot" style="animation-delay:.2s"></span>
<span class="thinking-dot" style="animation-delay:.4s"></span>
<span style="margin-left:4px;">Thinking...</span>
</div>
</div>
""", unsafe_allow_html=True)
# Input row
inp_col, btn_col = st.columns([5, 1])
with inp_col:
user_input = st.text_input(
"Ask", placeholder="Ask a question...",
label_visibility="collapsed", key="chat_input"
)
with btn_col:
send_btn = st.button("→", key="send_btn")
# ── Two-phase send ────────────────────────────────────────────────────
# Phase 1: user submits → store question, append to convo, rerun immediately
# so the question bubble appears before LLM is called.
# Phase 2: awaiting_answer is set → call LLM, append reply, rerun.
# Detect new submission
new_question = ""
if st.session_state.pending_input:
new_question = st.session_state.pending_input
st.session_state.pending_input = ""
elif send_btn and user_input.strip():
new_question = user_input.strip()
if new_question:
# Phase 1 — show question immediately
conversation.append({"role": "user", "content": new_question, "sources": []})
st.session_state.awaiting_answer = new_question
st.session_state.processing_answer = True
st.session_state.should_clear_input = True # Flag to clear on next rerun
st.rerun()
# Phase 2 — question is visible, now generate the answer
if st.session_state.awaiting_answer and st.session_state.processing_answer:
question = st.session_state.awaiting_answer
history_for_api = [
{"role": t["role"], "content": t["content"]}
for t in conversation
if not (t["role"] == "user" and t["content"] == question and t == conversation[-1])
]
try:
keyword_index = video.get("keyword_index")
reply, sources = chat_with_video(
question, history_for_api,
index, chunks, video_id, get_client(),
keyword_index=keyword_index,
)
conversation.append({"role": "assistant", "content": reply, "sources": sources})
except Exception as e:
conversation.append({
"role": "assistant",
"content": f"Sorry, something went wrong: {e}",
"sources": []
})
finally:
st.session_state.awaiting_answer = ""
st.session_state.processing_answer = False
st.rerun()
# Disclaimer
st.markdown("""
<div style="text-align:center;font-size:0.67rem;color:#555;padding:4px 0 2px;">
AI can make mistakes, so double-check it.
</div>
""", unsafe_allow_html=True)
# Clear chat
if conversation:
if st.button("Clear chat", key="clear_chat"):
st.session_state.conversations[video_id] = []
st.rerun()
Running the app
For the Running the app stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.
Prerequisites
For the Prerequisites stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.
Install dependencies
For the Install dependencies stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.
cd medium/chat-with-video
pip install -r requirements.txt
Run the app
For the Run the app stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.
streamlit run app.py
What are the main topics discussed?
What is mentioned about the relation of America with France?
Next improvements
For the Next improvements stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.
Conclusion
For the Conclusion stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.
References
For the References stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.
Operational checklist
For the Operational checklist stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state.
Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.