Repository navigation
Releases
0.8.0
A release about the first five minutes and about trust: a setup checklist, sample notes to try it on, a problem report you can read before it goes anywhere, other local servers besides Ollama, vaults of 50,000 notes, and tests that run in a real Obsidian.
Getting started
- Setup checklist. The first time you open the chat (and from the command or settings): is the server reachable, is there a chat model that can call tools, an embedding model, an index, each with a fix button. Missing models are downloaded through Ollama with progress; the recommended model follows the machine you pick (24 GB, 12–16 GB, 8 GB GPU, or none).
- Try it on sample notes. Creates a small folder of five linked notes and offers three questions, so you see a cited answer within a minute of installing.
Report a problem
- A command, a settings button, and a link on every error. It writes the GitHub issue for you (versions, operating system, model names, settings, index statistics, the last 50 log lines) and shows it for you to read and edit. Copy or Open on GitHub; the plugin itself sends nothing.
- What never goes in: API keys and tokens, URL query strings, e-mail addresses, your extra instructions, the vault's path and your home folder (shown as
<vault>and~), note contents, and chat text unless you tick Include the last chat.
Other local servers
- LM Studio, llama.cpp server, Jan, vLLM through their OpenAI-compatible API (
/v1/chat/completionsstreamed with tool calls,/v1/embeddings): Settings → Provider. Ollama stays the default and is what the benchmark measures. The plugin's own API reports the provider inapi.ollama()as an added field.
Languages
- Answers in your language: the language of the question by default, or one you set. The chat, settings, command names and the main dialogs read their strings from one dictionary (English complete), so a translation is one more file; most notices and the older dialogs are not moved yet.
Large vaults
- The index is stored in shards and built without quadratic work. On a generated 10,000-note vault the plugin's own share of the first index went from 134 s to 1.6 s, and editing one note rewrites 4 MB instead of 125 MB; a 50,000-note vault, which 0.7.0 could not have saved at all, indexes with 10 s of overhead and answers a semantic query in 0.28 s. Existing indexes are moved over without re-embedding.
- The vault outline and the Vault suggestions report no longer slow down with the square of the vault (backlinks from one reverse map; mentions by a name's rarest word): suggestions on 10,000 notes take under a second.
- Numbers, method and what they do not cover (the embedding model's own time, Obsidian's caches) are in the README under Large vaults.
Reliability
- A write is one whole note or nothing. Approving a diff after pressing Stop writes nothing; a note that changed while its diff was open is not overwritten; a failed write is not recorded and leaves the note as it was.
- A message that fails or is stopped part-way says how many changes it had already applied and offers to undo them all.
Quality
- Lint is at zero warnings with the Obsidian review rules: settings use the declarative API (so they show up in Obsidian's settings search on 1.13+), plain requests go through
requestUrl, and only the streamed replies usefetch. - End-to-end tests in a real Obsidian (wdio-obsidian-service, a scripted stand-in for Ollama): loading and the plugin API, chat → diff → approve → note changed, reject, undo, the change history, the problem report, the setup checklist, the sample notes and the settings page. They run in GitHub Actions on every change; the workflow fails unless all nine pass.
- Keyboard and screen-reader work: labels on icon buttons and inputs, a live region for the conversation, visible focus rings, a focusable diff, larger touch targets on phones.
- Beta channel for BRAT testers: pre-releases from
manifest-beta.json; the store'smanifest.jsonis untouched by a beta. - One feedback question, once, after the fifth answer; its buttons only open a GitHub page.
Reviewed before release
- The release was read by independent reviewers, and each finding was checked twice and then fixed: about 25 in all. Among them: after changing the embedding model, an interrupted rebuild could leave old vectors mixed into the index; every start re-read and rewrote the whole index; one unreadable index file discarded the others; an API key for an OpenAI-compatible server was still sent after switching back to Ollama; the problem report could carry a path or credentials in its issue title and listed private folder names from settings; the one-time survey came back after every answer; and existing users would have been shown the first-run checklist.
Benchmark
- The full 24-scenario accuracy benchmark was not repeated for this release. It was started on a laptop without a discrete GPU, where each model call takes about 30 seconds, and stopped after the first seven scenarios: 41 of 41 checks with
qwen3.5:4band 40 of 41 withqwen3.5:9b(the miss: one tool call more than allowed on the lecture-by-concept question). The numbers in the README's benchmark table are still the 0.5.3 runs. - New: a large-vault benchmark (
npm run bench:scale), whose results are in the README under Large vaults and inbench/0.8.0-scale.json.
Not verified for this release, said plainly: nothing was opened in Obsidian on a desktop by hand. The end-to-end tests cover the chat, diff, undo, history and report dialogs, the setup checklist, the sample notes and the settings page in Obsidian on Linux in CI; the model download, the survey and everything from 0.7.0 that needed Obsidian (mic, scanned PDFs, scheduled jobs on their timer) have unit tests for their logic but were not exercised in a running Obsidian. The OpenAI-compatible path is tested against the wire format, not against LM Studio, llama.cpp, Jan or vLLM themselves. Accessibility was reviewed in code, not with a screen reader, and nothing was tried on a phone. The large-vault numbers are from the plugin's core under Node with a stand-in embedding model, not from a 50,000-note vault open in Obsidian.
0.7.0
Answers link the exact paragraph, the agent reads Lecture Companion's lectures and your images, quizzes you, links concepts across courses, runs scheduled jobs that wait for your approval, and keeps a change history you can undo from.
Citing the passage, not just the note
- A Passages line under answers from notes links the paragraphs the answer used: those it quotes, the transcript paragraphs whose
[m:ss]it names, or ones it closely paraphrases. A paragraph with a block id is linked as[[note#^id]]. - Block ids are only written with your approval. A paragraph without one gets an id proposed (
^t21-29for the transcript paragraph at 21:29, a short hash otherwise); all of one answer's ids for a note are one diff. Reject it and the link points at the paragraph's heading ([[note#Transcript|note › 21:29]]), which needs no write. Settings → Paragraph citations: block ids, headings only, or off.
Lecture Companion
- Lectures and transcripts as data. With Lecture Companion installed (its API version 1), the agent gets
list_lecturesandlecture_transcript: the matching transcript paragraphs with their[m:ss], instead of searching notes and following conventions inAGENT.md. Without it everything works as before. - Voice input. A mic button in the chat records, converts to 16 kHz mono WAV and transcribes through Lecture Companion's whisper server into the message box. Without Lecture Companion the button is not shown.
- One Ollama for both. Use Lecture Companion's Ollama settings in settings copies its URL and model.
- This plugin's own API,
app.plugins.plugins["local-vault-agent"].api(version 1):openChatAbout(path)andollama(). Both APIs are documented in the README.
Studying
- Quiz me. "Quiz me on CSCI 470 since October 1" (or the command, or
/quiz): questions written from those notes, asked one at a time, each answer graded against its source paragraph with the answer and a link to it; missed ones go to a review note after you approve the diff. - Suggest concept links for this note. Concept or topic notes the open note mentions (or is close to by meaning) become links in one diff, optionally recorded as
parent/childlink properties for Branching Table of Contents, so one concept gathers lectures from every course.
Images and scans
read_imagereads an image (text, handwriting, diagrams, charts) or a scanned PDF page with a local vision model;read_notetells the model which images a note embeds andread_filewhich PDF pages are scans. Readings are cached by content hash. Settings → Vision model: empty uses the chat model when it can see (qwen3.5can), or set e.g.qwen2.5vl:7b; without one the tool is not offered.
Scheduled jobs
- Nightly Process my daily notes and a weekly digest note, off by default. They run when Obsidian is open at or after their time and never write: what they would change waits until you are back (a notice and a status-bar badge), then each change is a diff you approve. A note you edited meanwhile gets the change carried over when the edits do not overlap, and is skipped when they do.
Change history
- Every approved change is kept (the last 300, across restarts) with its time, chat and diff. Show change history (
/history) lists them with an Undo button. - Undo keeps later edits. If a note was edited after the change, only the change is taken back; if the later edits overlap it, the undo is refused instead of overwriting them. Deleting a note the agent created is refused once you have written in it. An undo can itself be undone.
Benchmark
- Four new scenarios and a new check (Lecture Companion tools, quiz, the nightly job in an overlay, reading an embedded image; a passage link on the transcript question): 32 of 34 checks with
qwen3.5:9b. The misses are the model's: a queued topic note without its Sources section, and a drawn "47" read as "41". The full 24-scenario run was not repeated for this release. - The benchmark harness builds again: it still imported the Vault Suggestions note writer that 0.6.0 removed.
Only checkable inside Obsidian (built and unit-tested, not exercised in a running Obsidian for this release): the mic button and recording (microphone permission, MediaRecorder decoding), rendering scanned PDF pages for the vision model, the change-history and concept-link dialogs, the status-bar badge and the reminder notice, scheduled jobs firing on their timer, and the Lecture Companion API calls against the real plugin (its side was built in parallel to the same interface; here it was exercised through a fake that follows the specification).
0.6.0
Vault suggestions move into the chat and stop needing the model, and a new message no longer re-reads the whole prompt.
Vault suggestions (/garden)
- Nothing is written to the vault. The report is a chat message with plain bullets; the Vault Suggestions checklist note is gone, because its open checkboxes showed up in every task query over the vault. If an earlier version left that note behind, the report offers to move it to the trash.
- The mechanical fixes run without the model. Repoint moved links, Link mentions and Remove identical copies are worked out by the plugin: the same diff dialog as an agent edit (with Approve all), done in about a second, undone with
/undo-all. Asking a local 9B model to do this took ten minutes of reading notes it did not need to read. Only Review similar notes and Suggest links for orphans still ask the model, and only for a proposal. - A broken link says whether its target moved.
[[Old folder/Note]]is reported with where the note is now when exactly one note or file has that name; those are the links the button repoints. - Fewer false findings. A mention has to be a distinctive name: several words, or one word that could not be an ordinary one (
GenKIC), with capitalised names matched as written; never a note the page already links to, or the note its own folder is named after. Duplicates are identical copies (found without the index; frontmatter and the folder inside a wikilink do not count as a difference) or notes that share at least half their lines, not everything the embedding index calls similar, which included every lecture built on the same template and slide deck. A note that is one query block or one embed is not a stub, and the conventions note is not an orphan. - Prompts sent by a button no longer attach every note they name, which had added tens of thousands of characters the model then read a second time.
Speed
- A new message starts from the cached prompt. The date, the open note and today's daily note moved out of the system prompt into a line sent with your message. Ollama reuses what it has already read only up to the first difference, and the system prompt comes first, so opening another note between two messages used to mean re-reading the outline, the tool list and the conversation. On a laptop with an integrated GPU (Ryzen 7940HS, Radeon 780M,
qwen3.5:9b) the wait before the first token of a second message went from 58 s to 1 s. - Only changed chunks are embedded again. A note that grows a paragraph at a time (a lecture being transcribed, a daily note being typed) had every chunk re-embedded on each save; now only new or edited chunks go to the embedding model, and an edit is picked up after ten seconds of quiet instead of two.
- README: what to expect without a discrete GPU, and why Keep model loaded should be long there.
0.5.3
Faster turns and a far shorter first index, with the same benchmark score.
Why it was slow. Two things compounded. The history the agent carried per step grew with the context window, so at a 96k window a long session reached 60k tokens; and once over budget the trimmer reshaped older messages on every step, which invalidated Ollama's prefix cache, so each step re-read the whole conversation. On a 12 GB card a 9B model at 96k also leaves no room for the embedding model, so Ollama evicted one to load the other on every search, reloading the chat model each time.
Changed
- History is capped at the 32k budget whatever
num_ctxis set to (the benchmark scored the same at 32k and 64k), and finished turns are compacted once when a new turn starts instead of being reshaped on every step: reasoning dropped, old tool results cut to 400 characters, oldest turns removed. Within a turn the messages sent to Ollama are append-only, so the prefix cache holds and each step prefills only what is new. - A context window above 32k is brought back to 32k once by the settings migration (set it again if you have the VRAM); the setting's description now explains the trade-off.
- Gists are off by default. The pass runs the chat model once per note; on this card that was 230 s for the benchmark vault and bought one check, on a laptop GPU with a large vault it is the better part of an hour. The embedding index that search depends on is unchanged and stays on. Existing installs with gists on are switched off once; the toggle turns them back on.
- Embedding batches of 64 passages instead of 16 (same speed per passage on a GPU, a quarter of the round trips on a CPU-only machine).
- The plugin tells you when the model is spilling into system RAM. After a turn it checks Ollama's
/api/ps; if the chat model is only partly in VRAM, one Notice says how much and what to do (a smaller window orqwen3.5:4b). - README: "Settings worth knowing" rewritten around these findings, plus an "If it feels slow" checklist.
0.5.2
An installed embedding model is recognised as installed.
Fixed
- "model not installed: nomic-embed-text" with the model installed. Ollama lists a model pulled without a tag as
nomic-embed-text:latest, and the plugin compared the names literally, so the startup Notice, the settings Test connection button, the model picker in the panel and the tuning advice all called the model missing (and the picker showed a spurious "(not installed)" entry). Names are now compared the way Ollama resolves them: case-insensitively and with:latestoptional.ollama pull nomic-embed-textandnomic-embed-text:latestin settings are the same model.
0.5.1
The clean-up that used to fail now passes every run, a rejected change really stops the agent, and the benchmark grew to 20 scenarios.
Fixed
- A rejected change ends the turn properly. After you reject a diff the agent refuses every further write in that turn, tells you what was rejected and asks what you want instead. Before, the model sometimes said "Done" after a rejection or retried a variant of the same edit.
- Surveys the model could misread.
list_sectionswithcontaining: ["## Capture"]andheading: "Capture"returned nothing (the heading line is not part of a section's body), which the model took as "no Capture sections exist". The text is now treated as the heading, and an empty match says "N sections exist, but none contains X". Sections that contain every requested text are marked→ allwith a line telling the model to act on them without reading them first; the 9B model used to re-read all 40 sections to "check". - Clean-ups across many notes no longer ask for one giant
edit_notescall: at most 8 edits per call, one call per note. An 18-edit call made the model emit nothing at all. - Process my daily notes requires the Sources section in every topic note it writes.
- An attached note that only half answers a question is followed up with a search instead of "the note does not mention it".
- The Sources footer also covers answers taken straight from
search_passagesresults, naming the hits the answer mentions. - Ask your lectures. Lecture Companion now writes its note conventions into
AGENT.md(a block between markers), and the benchmark vault's lecture sections carry a### Transcriptof**[m:ss]**paragraphs in that shape; the newlecture-timestampscenario ("what did the professor say about contamination in the isolation forest lecture, and at what time?") asks for the fact, the timestamp and the note link: 7 of 7 in three runs once the Sources footer learned to name the note whose passage the answer quotes rather than a note that merely shares the subject's name (the paperResearch/Isolation Foresthad been credited for a transcript quote). 20 scenarios, 122 checks.
Measured
- 20 scenarios, 122 checks (new: a lecture question answered with the transcript timestamp, a follow-up turn that must use the conversation, a question with no answer in the vault with a check for invented links, a rejected change). Final build, three runs of the defaults on all 20 scenarios: 121, 120, 120 of 122 in 167–197 s (113, 113, 112 of the earlier 115 checks) (the 0.5.0 code scored 87–92 of the old 96 checks on the same night); the clean-up now surveys once and edits once in most runs (it passed 0 of 3 before). Model time is 90–97% of wall time.
qwen3.5:4bscores 108 in under two minutes and is now the recommended 8 GB model. 64k context and thinking on did not beat the defaults. Table and reasoning in the README. - The usage callback carries Ollama's prompt, generation and load times; the benchmark reports model seconds and steps per scenario.
- Vault size: the benchmark on a 3× vault (997 notes,
SCALE=3in the generator) scored 110 of 115 in 214 s with 83 tool calls against 112, 155 s and 66 calls at 339 notes; the index took 47 s instead of 8–10 s and the system prompt grew from ~5.0k to ~6.2k tokens. The outline had fallen back to an alphabetical prefix that droppedSchool/,Projects/andPeople/from the prompt; folders that do not fit now collapse to one line with their count and name range, after which the 3× run scored 114 of 115, and three more runs on the 3× vault scored 110, 114 and 115 of 115 (the first perfect run of the night) in 162–307 s.
0.5.0
Four new things to do with the agent, measured benchmarks, and a model that no longer talks itself into loops.
New
- Process my daily notes. One command (or
/daily) that surveys the inbox section of every daily note, compiles the substantial ones into topic notes with links back to the source days, collects code blocks into a snippets note, and replaces each compiled inbox with- Moved to [[Topic]]so the daily notes stay timelines. Options dialog for the folder, heading, destination, look-back and whether to move; every write is diff-approved. - Edit the selected text with an instruction. Select a passage, run the command (or
/edit), type "make this a table" or "past tense", approve the diff. The agent edits exactly the passage you selected. - Suggest vault improvements (
/garden). A deterministic pass, no model needed, that writes a Vault Suggestions checklist note: broken links, daily inboxes that filled up, notes mentioned by name without a link, near-duplicate notes (from the semantic index), orphans and stubs. Each section says what to ask the agent to fix it. - Test connection and speed. The settings button now also runs a short generation at your context size, reads how much of the model is on the GPU versus in RAM, and gives one-line advice with buttons: use a larger context when there is room, a smaller one or a smaller installed model when it is spilling. Flags a chosen model that cannot call tools.
replace_sectiontool: rewrite or empty everything under a heading without copying its old text.list_sectionssurveys one heading across notes:heading: "Capture"lists that section of every daily note with its size and first line, so "which daily notes have a full inbox" is one call instead of 120.
Measured
npm run e2e:largeis now a scored benchmark: 14 scenarios on a 339-note synthetic vault (questions, enumeration, clean-ups, moves, the daily-notes job, and the loop failures seen in a real vault), each timed and counted for tool calls, errors, guessed paths and repeated calls; results are saved tobench/<version>.json. The README carries the table.
Changed
- Thinking is off by default. Across the benchmark matrix,
qwen3.5:9bwith thinking off passed as many checks as with thinking on (92–94 of 96 against 90–91) in a third of the wall time and with fewer tool calls; the agent loop already gives the model one step at a time to reason over. Users still on the old "Auto" default are moved to "Off"; "On" chosen explicitly stays. - A reply that ends by announcing a step ("To find the loss, I'll check the NNUE Training note.") without taking it is sent back once with "do it now". Models answering without thinking did this on about one question in sixteen.
Fixed
grep_vaultwithfolder: "Daily"searched onlyTemplates/Daily.md: a folder name that also resolved to a note was taken as that note. Any vault with a daily-note template was affected; the folder now wins whenever it exists. Found by the benchmark's blockquote scenario.- An answer given straight from a
grep_vaultresult ("the isolation-forest lecture was on 2026-11-06") now gets the Sources footer naming the matched notes, like answers from notes that were read. - Sampling: temperature 0.6 with top_p 0.9, top_k 20 and a small presence penalty, per the Qwen3 model card, which warns that near-greedy decoding causes endless repetition. Users still on the old 0.3 default are moved; a changed value stays.
- The system prompt is built once per turn instead of once per step, so Ollama's prefix cache is hit on every step of a long task even after a write (before, every write invalidated it).
grep_vaultalso recognises escaped metacharacters (!\[\[Recording) as a regex sent by mistake withregex: false.- When a long task fills the window and an earlier tool result is shortened with "call the tool again if you still need it", calling it again is allowed; 0.4.1's repeated-call guard refused it and could end the turn early. Found by the benchmark's clean-up scenario.
- After a write, the guard forgets only reads of the changed note and vault-wide results (grep, lists, searches); re-reading an untouched note "to verify" is refused with the earlier result, which is where a 9B model spent half of a clean-up's time.
0.4.1
The agent no longer gets stuck. Seen in a real session on a 9B model: one grep_vault call repeated 60 times with identical arguments, 150 read_note calls guessing every calendar date in Daily Notes/ (60 % of them misses), and a regular expression sent as literal text that found nothing and turned into a false "there are no code snippets to move".
Reliability
- Repeated-call guard: a read-only tool called again with exactly the same arguments in a turn is not run; the model is shown the earlier result and told to do something else. At the third repeat the turn ends with a wrap-up answer instead of burning the whole call budget. The guard forgets everything after a write, so a re-read after an edit sees the new content.
grep_vaultrecognises a regular expression sent withregex: false(^\s*>,a|b,[A-Z]): when the literal search finds nothing it searches as a regex and says so. Spaces around|are trimmed.- A path that does not exist now returns the notes that do exist in that folder (all of them for a small folder, the nearest by name for a large one), or the folders that exist and similarly named notes, so one miss teaches the model the folder instead of inviting the next guess.
list_notesshows each note's size, so empty template notes can be skipped without reading them.- Prompt: never guess paths; survey one section across many notes with a single
grep_vaultcall with context; a search that finds nothing is the answer for that pattern; "messy", "empty" or "long" are structural questions forlist_notes,list_sectionsandgrep_vault, not for semantic search.
0.4.0
Reliability on big vaults, and the agent behaves more like a coding agent.
Reliability
- Context guards: when a long task fills the window, older tool results from the same turn are shortened (the system prompt and your request are never pushed out), a single batch of tool calls is capped, and a blank reply is nudged once. Found by running a 330-note synthetic vault.
grep_vault: grep across every note, each match tagged with its enclosing heading, optional context lines.list_sections: per-section yes/no flags for up to four texts ("which lecture sections have a recording but no Summary") without reading anything.edit_notes: many edits across many notes in one call, one approval per note; items can address a line byline_containswithin aheading. A clean-up across 60 lecture sections is now two tool calls (measured: 17/18 embeds removed, none wrongly, 115 s on a 9B model; before: 0/18 and a full context window).- At most 8 tool calls run per step; the rest are refused with a hint to grep instead of reading everything. The prompt tells the model to use it instead of reading every note to find something.
- Outline lines name the file explicitly, and tools tolerate a path copied with outline decorations.
- Backticked note paths in replies become clickable wikilinks.
Speed
- Re-reading an unchanged note within a turn is short-circuited.
- Readable tool cards ("Searched passages for …", "8 results") and a live "3 tool calls · 42 s" status while a turn runs.
Capabilities
read_fileextracts the text of PDFs (Obsidian's pdf.js) and text attachments.list_notesfilters bymodified_after/modified_before.- Today's daily note (per the Daily Notes core plugin settings) is named in the system prompt.
- README documents measured context/VRAM/speed numbers for 32k–128k windows with q8_0 KV cache.
Polish
- Approval dialog: Enter approves, A approves all in this chat, Esc rejects.
- The chat is saved and restored across panel closes and restarts; New chat clears it.
- One notice per session when Ollama or the chosen models are not available.
- Default max tool calls per message raised to 60.
0.3.1
Two fixes found in live testing right after 0.3.0:
- Small models copied whole vault-outline lines ("Folder/Note — heading › heading") as note paths and then failed to find the note. Outline lines now name the file explicitly (
Note.md — …), the prompt says the path stops at the dash, and every path-taking tool strips that decoration before giving up. - Replies that cited notes as backticked paths (
Projects/Note.md) are rewritten to clickable wikilinks ([[Projects/Note]]) when the path resolves to a note.