iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
You do not need a vector index to give an AI model useful context from your own notes. A workflow described in September 2026 does it with plain Markdown files, the ripgrep command-line search tool, and a person who decides which lines go into the prompt. It is a retrieval choice with a clear strength and a clear ceiling, and the decision depends on how your notes are written and how you search them.
The account comes from a Spanish-language write-up by Donweb, dated 16 September 2026, of a Dev.to post titled “My notes became a personal RAG with no embeddings.” The original post was not available for direct review, so the workflow details, the figures, and the usage claims below are the Donweb account’s description. None of them has been independently verified or tested under controlled conditions. The Donweb account is at https://blog.donweb.com/rag-sin-embeddings-notas-personales-grep-2/.
What retrieval-augmented generation means in this case
Retrieval-augmented generation (RAG) means finding material you already have and supplying it to a model as context for its answer. The usual picture involves vector embeddings: text is converted into numerical vectors, stored in an index, and ranked by similarity to the question. Embeddings are one retrieval method. Literal text search followed by human review is another, and it can supply the same kind of context. The reported workflow uses the second method. Whether that counts as “RAG” depends on your definition, since some definitions assume automated ranking. The useful point is practical: the model receives retrieved text either way.
Free tools Windows power users keep installed
One-click scans. No signup required.
The workflow as described
The Donweb account describes a short loop built on personal Markdown notes and a terminal search:
#1 Best Overall
- Write notes as plain text. The notes are ordinary Markdown files in a folder, with no database or metadata layer.
- Search for the words you expect you used. The account’s example command is
rg -i --sort path 'palabra' notes/. The-iflag makes the match case-insensitive,--sort pathorders results by file path, andnotes/limits the search to that folder. The example term,palabra, is Spanish for “word”; you would substitute a term from your own notes. - Read the matching lines. The person scans the returned lines and the files they come from.
- Select the passages that matter. The person, not a ranking pipeline, decides what is relevant.
- Paste the selected passages into the prompt alongside the question you want the model to answer.
The account reports eight months of use and a collection of a few thousand lines of notes. Both figures come from that one reported case; they are not measurements of how the method performs across users or note collections.
What the method buys
The account presents three advantages, and each is easy to check against your own situation.
- Nothing to maintain beyond the notes. There is no embedding model to choose, no index to rebuild after edits, and no retrieval service to keep running.
- A person reviews every candidate. Before a passage reaches the model, someone has read it in context. That review catches stale notes, off-topic matches, and passages that match the word but not the meaning.
- The reason for each result is visible. A line appears because it contains a specific term. You can see why it was retrieved, which is harder with similarity scores.
Where literal search breaks
Literal search only finds text that shares wording with the query. Two failure patterns follow from that.
- Vocabulary mismatch. A note about “retry with backoff” will not appear for a search on “resilience.” A note that says “deadline” will not appear for “timeout.” If you wrote the note in one language or phrasing and search in another, the match can disappear entirely.
- Silent misses. A note that never matches your query is not shown to you, so you usually do not know it exists. The method is only as good as the terms you think of, and the failure is invisible unless you later stumble on the missing note.
The account does not measure how often either failure occurs. Treat it as a risk to check in your own notes, not as a rate.
Comparing literal search with semantic retrieval
The useful comparison is between two retrieval methods on the same five axes. The account does not benchmark semantic retrieval, so the right-hand column describes how such layers generally behave, and cells that the account does not address are marked accordingly.
| Axis | Literal search with manual selection | Semantic retrieval layer |
|---|---|---|
| Query vocabulary | Needs the words used in the note, or close variants. | Can match related concepts expressed in different words, with quality depending on the embedding model and index. Not evaluated in the Donweb account. |
| Corpus scale and review effort | Review is quick while result sets stay small enough to scan. The account gives no threshold. | Returns ranked candidates that still need checking; the size at which review becomes burdensome is not stated. |
| Who searches | Works well for the author, who knows the vocabulary. Harder for someone who did not write the notes. | Can reduce dependence on the author’s exact wording, though the account does not test this. Not stated. |
| Human oversight | Built into the workflow: the person reads each candidate before use. | Oversight is optional and depends on how the system is designed. Not stated in the account. |
| Maintenance | Only the notes and a search command. Not stated as requiring other components. | Requires an embedding model, an index, and updates when notes change. Costs are not quantified in the account. |
When to add semantic retrieval
The account treats these situations as reasonable triggers for adding semantic search to a personal knowledge base. They are decision criteria it reports, not thresholds established by testing.
- You repeatedly know a note exists but cannot recall its wording.
- Related ideas in your notes use different vocabulary, so one search term keeps missing relevant passages.
- The collection has grown to the point where scanning matches takes more time than the answer is worth.
- Other people need to search notes they did not write and cannot be expected to know the author’s terms.
If none of these apply, literal search is likely enough. If one applies, the simplest next step is to confirm that the misses are real before adding infrastructure.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Adding a semantic fallback only when misses justify it
The most defensible way to combine the two methods is to keep literal search as the first step and add a semantic fallback only after real missed-recall cases appear. Keep a short log of questions where you knew the answer existed in your notes but the search did not surface it. If that log stays short, the extra system is hard to justify. If it grows, you have evidence for the change.
The Donweb account describes a semantic fallback as the author’s contemplated next step. That fallback was not shown to be built or effective in the account, so treat it as a plan rather than a result.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.A minimal setup to try
- Install ripgrep using your system’s package manager, for example
brew install ripgrepon macOS orsudo apt install ripgrepon Debian and Ubuntu. The account does not establish a setup for other systems. - Create a notes folder such as
notes/and save each entry as a dated Markdown file, for example2026-10-09-retry-policy.md, so results sort by date in the file name. - Search with several term variants. Run
rg -i 'retry' notes/, thenrg -i 'backoff' notes/, and note which variant finds the passage you want. - Review before pasting. Copy only the lines that answer your question, plus enough surrounding context to be understood, into the AI conversation.
- Record misses. When a relevant note fails to appear under any term you try, add the question and the note’s wording to a log.
Source provenance
The Donweb account links to a Dev.to post in the same publishing series, which states that its posts are written and published autonomously by an AI agent working from the Simple Memo project’s public records. That statement applies to that post. It does not establish who wrote the target post or what evidence sits behind its figures. The series post is at https://dev.to/simple_memo/a-null-return-hid-three-billed-runs-inside-an-automation-ledger-jmg.
Recommendation
Start with the simplest retrieval method that works for your actual notes. For a small, familiar collection that you wrote yourself, plain-text files and a literal search with manual selection are a reasonable place to begin, and the Donweb account reports they worked for one author over eight months. Add semantic retrieval when your own logged misses show that vocabulary, scale, or other people’s searches are causing failures you can measure. This recommendation is a practical synthesis of the cited account, not a finding from a controlled comparison.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
