Summary and Next Steps

What you built, what you learned, and the most interesting directions to take it next

You started this course with a vault full of notes you could not find, and you finished it with a tool that answers questions about them and shows its working. Nicely done.

What you built

Ask My Vault - eight Python modules, three dependencies, and a command you can run from anywhere:

amv index                                        # keep it current
amv search "why did the motors stall"            # see the raw matches
amv ask "why did the motors stall" --tag robots  # get a cited answer
amv stats                                        # check on the index

What you actually learned

The tool is nice. The transferable knowledge is nicer, because it applies to any RAG system you build after this one.

Embeddings turn meaning into geometry. Similar text lands at nearby points, and “search” becomes “find the closest points”. Once that clicks, the rest is engineering.

Chunking is where the quality lives. More than the model, more than the prompt. A chunk that carries its own context can be found; one that does not, cannot.

Retrieval and generation fail differently, so debug them separately. The search command exists purely so you can tell in ten seconds which half is broken.

Grounding is a prompt discipline, not a feature. “Use only these sources”, “cite them”, “never invent a fact”, and refusing to call the model with nothing retrieved. Four rules that are the difference between a research assistant and a confident liar.

Filters run before ranking, so they change what wins - not just what survives. That is what makes metadata worth storing.

If you do not measure it, you are guessing. Twenty questions in a YAML file turn “that felt better” into a number you can compare.

Where to take it next

Roughly in order of how much fun they are relative to the effort.

A better embedding model. Swap the default for something larger through Chroma’s embedding_function argument. Measure it with your evaluation set before and after - you may be surprised how little difference it makes compared to fixing chunking.

A re-ranker. Retrieve 20 candidates, score them with a cross-encoder, keep the best 5. This is the standard next step in production RAG and it works, at the cost of another model and real latency.

Follow the graph. You already store links on every chunk. When a note is retrieved, pull in the notes it links to. Obsidian users build dense link graphs and there is a lot of signal in there you are currently ignoring.

Conversation. Keep the message history and let follow-up questions work. “What about the quad version?” needs the previous turn to make any sense. You will need to rewrite follow-ups into standalone questions before retrieving - that is the interesting part of the problem.

Fully offline. Swap the Claude call for a local model through Ollama and the whole pipeline runs with no network at all. Quality drops, privacy goes to absolute, and on a Raspberry Pi 5 with a small model it is genuinely usable.

An Obsidian plugin. Run the retrieval as a small local HTTP service and call it from a plugin, so you can ask questions without leaving Obsidian. Our FastAPI course covers the server side.

Point it at something else entirely. Nothing in the pipeline is Obsidian-specific except the frontmatter and wikilink parsing. Swap vault.py and you have a RAG over your code documentation, your email archive, or a folder of PDFs.

What RAG cannot do

Worth restating, because it is easy to forget once the tool is working well.

It cannot know what you did not write down. RAG searches your notes. If the decision only ever lived in your head, no amount of retrieval finds it.

It cannot tell that your notes are wrong. A confidently cited answer sourced from a note containing a mistake is a confidently cited mistake. The citation tells you where a claim came from, not that it is true. Click through.

It cannot do maths across your whole vault. “How many robots have I built?” needs counting, not similarity. Retrieval finds relevant text, and no set of six chunks contains a count over a thousand notes.

It cannot replace a good note. The best thing you can do for your RAG is write clearer notes with real headings. Everything in this course amplifies the quality of your notes - it does not create it.

Prerequisites and related courses

If you want to go deeper on the pieces:

• Obsidian - the note-taking side, if your vault needs some love
• Learn Python - the fundamentals
• FastAPI - for wrapping this in a local service
• SQLite - Chroma is SQLite underneath, and knowing what is in there helps
• Docker - if you want to run the indexer on a Pi as a container

One last thing

Run amv ask on a question you have genuinely wondered about for months. Something you know you wrote down and never found again.

That moment - when it comes back with the answer and a link to a note you wrote two years ago and completely forgot - is the whole point of building this.

Thanks for following along. Go and ask your vault something.

< Previous

You can use the arrows  ← → on your keyboard to navigate between lessons.