Uploaded September 2026 | Updated November 2026, 11 hours ago
Learn the ingestion half of a RAG system in LangChain: how to chunk documents, turn them into embeddings, store them in a Qdrant vector database, and search them by meaning instead of keywords.
Retrieval can only be as good as what you ingested, so this lesson stays entirely on ingestion and leaves the production build for Day 10. You will see why a 300 page PDF cannot go into a prompt, how chunk size and overlap change what comes back, what an embedding really is, and why cosine similarity finds "baggage" when you asked about "luggage". Then we run Qdrant in Docker, split a real file with RecursiveCharacterTextSplitter, embed every chunk, ingest the documents with metadata, and query them with similarity_search and metadata filters. The lesson ends on the problem behind most beginner RAG hallucinations: similarity search always hands back a chunk, even when the answer was never in your database. Day 10 picks up right there and fixes it.
๐บ Full playlist: youtube.com/playlist?list=PLW4pPr9JCovI
โฑ Chapters:
0:00 What Day 9 covers and why ingestion comes first
2:36 Why a RAG system needs document chunking
5:02 Why LLM accuracy drops as context fills up
6:10 Chunk size and the 2000 token embedding limit
7:31 The 250 to 800 token sweet spot
9:29 Chunk overlap and why 10 to 20 percent
10:25 What an embedding actually is
11:57 Projecting text into vector space
13:31 Why similarity search beats keyword search
16:11 When similarity search returns the wrong chunk
17:10 Keyword match plus vectors is hybrid search
17:52 Choosing a vector database and why Qdrant
19:07 Running Qdrant in Docker
22:01 The full RAG pipeline diagram
23:53 Sample data and loading the first file
25:32 RecursiveCharacterTextSplitter, chunk_size and chunk_overlap
28:33 Why chunks are not exactly 250 tokens
29:37 The separator order the splitter tries
31:49 Comparing chunk sizes 100, 250 and 1000
34:18 Turning text into vectors with an embedding model
36:01 How cosine similarity works
37:06 Dot product of two vectors in code
38:52 Luggage vs baggage scores 74 percent
40:36 Embedding every chunk
41:25 Ranking chunks against a real question
44:50 Connecting to the Qdrant client
45:20 Turning chunks into Document objects with metadata
47:12 Ingesting into the day9_baggage collection
49:33 similarity_search from the vector store
50:54 Filtering by metadata before searching
53:11 The refund question with no matching data
55:19 What Day 10 covers
๐ Resources:
Code, notebook and docker-compose: github.com/laxmimerit/Langchain-10-Days-Bootcamp
Qdrant local quickstart with Docker: https://qdrant.tech/documentation/quickstart/
LangChain text splitters: docs.langchain.com/oss/python/integrations/splitters/index
LangChain vector stores: docs.langchain.com/oss/python/integrations/vectorstores/index
Embedding models on Hugging Face: huggingface.co/models?pipeline_tag=sentence-similarity
Free models on OpenRouter: openrouter.ai/models
LangChain install guide: docs.langchain.com/oss/python/langchain/install
UV package manager setup: youtube.com/watch?v=pcYaP1lGd3s&list=PLc2rvfiptPSQdX4jeLgW4nf5PId4j1msI&index=8
๐บ Watch next:
Tools and Tool Calling Explained: youtu.be/h4Stm0XrmKM
๐ Go deeper with my Udemy courses:
Master Langchain v1 and Ollama - Chatbot, RAG and AI Agents: kgptalkie.com/langchain
Advanced RAG: Build and Deploy Production GenAI Apps: kgptalkie.com/advanced-rag
If this made chunking and embeddings click, hit like so more beginners find it, and tell me in the comments which chunk size worked best on your own data. Subscribe and turn on the bell, Day 10 is the production grade RAG build and it starts exactly where this one stops.
#LangChain #RAG #VectorDatabase #Qdrant #GenAI
Learn the ingestion half of a RAG system in LangChain: how to chunk documents, turn them into embeddings, store them in a Qdrant vector database, and search them by meaning instead of keywords.
Retrieval can only be as good as what you ingested, so this lesson stays entirely on ingestion and leaves the production build for Day 10. You will see why a 300 page PDF cannot go into a prompt, how chunk size and overlap change what comes back, what an embedding really is, and why cosine similarity finds "baggage" when you asked about "luggage". Then we run Qdrant in Docker, split a real file with RecursiveCharacterTextSplitter, embed every chunk, ingest the documents with metadata, and query them with similarity_search and metadata filters. The lesson ends on the problem behind most beginner RAG hallucinations: similarity search always hands back a chunk, even when the answer was never in your database. Day 10 picks up right there and fixes it.
๐บ Full playlist: youtube.com/playlist?list=PLW4pPr9JCovI
โฑ Chapters:
0:00 What Day 9 covers and why ingestion comes first
2:36 Why a RAG system needs document chunking
5:02 Why LLM accuracy drops as context fills up
6:10 Chunk size and the 2000 token embedding limit
7:31 The 250 to 800 token sweet spot
9:29 Chunk overlap and why 10 to 20 percent
10:25 What an embedding actually is
11:57 Projecting text into vector space
13:31 Why similarity search beats keyword search
16:11 When similarity search returns the wrong chunk
17:10 Keyword match plus vectors is hybrid search
17:52 Choosing a vector database and why Qdrant
19:07 Running Qdrant in Docker
22:01 The full RAG pipeline diagram
23:53 Sample data and loading the first file
25:32 RecursiveCharacterTextSplitter, chunk_size and chunk_overlap
28:33 Why chunks are not exactly 250 tokens
29:37 The separator order the splitter tries
31:49 Comparing chunk sizes 100, 250 and 1000
34:18 Turning text into vectors with an embedding model
36:01 How cosine similarity works
37:06 Dot product of two vectors in code
38:52 Luggage vs baggage scores 74 percent
40:36 Embedding every chunk
41:25 Ranking chunks against a real question
44:50 Connecting to the Qdrant client
45:20 Turning chunks into Document objects with metadata
47:12 Ingesting into the day9_baggage collection
49:33 similarity_search from the vector store
50:54 Filtering by metadata before searching
53:11 The refund question with no matching data
55:19 What Day 10 covers
๐ Resources:
Code, notebook and docker-compose: github.com/laxmimerit/Langchain-10-Days-Bootcamp
Qdrant local quickstart with Docker: https://qdrant.tech/documentation/quickstart/
LangChain text splitters: docs.langchain.com/oss/python/integrations/splitters/index
LangChain vector stores: docs.langchain.com/oss/python/integrations/vectorstores/index
Embedding models on Hugging Face: huggingface.co/models?pipeline_tag=sentence-similarity
Free models on OpenRouter: openrouter.ai/models
LangChain install guide: docs.langchain.com/oss/python/langchain/install
UV package manager setup: youtube.com/watch?v=pcYaP1lGd3s&list=PLc2rvfiptPSQdX4jeLgW4nf5PId4j1msI&index=8
๐บ Watch next:
Tools and Tool Calling Explained: youtu.be/h4Stm0XrmKM
๐ Go deeper with my Udemy courses:
Master Langchain v1 and Ollama - Chatbot, RAG and AI Agents: kgptalkie.com/langchain
Advanced RAG: Build and Deploy Production GenAI Apps: kgptalkie.com/advanced-rag
If this made chunking and embeddings click, hit like so more beginners find it, and tell me in the comments which chunk size worked best on your own data. Subscribe and turn on the bell, Day 10 is the production grade RAG build and it starts exactly where this one stops.
#LangChain #RAG #VectorDatabase #Qdrant #GenAI










