More retrieved context can reduce answer quality
Retrieval quality depends on relevance, not simply context volume.
[[Resources/AI/Retrieval precision]][[Resources/AI/Citation coverage]]Checked August 2026
Local knowledge workbench
Build an AI second brain in Obsidian—without hiding the files
Building a Second Brain is a knowledge-work method; Obsidian is one local Markdown substrate. This workbench turns a source packet into proposed concept notes, preserves exact provenance, commits only reviewed changes, and updates links and backlinks as you edit the graph.
Then ask a question through keyword, semantic, graph-neighbor, or hybrid retrieval. The exact note chunks and prompt envelope stay visible, as do stale indexes, private-folder exclusions, injected source text, write approval, and every path that can leave the device.
- Manipulate
- Source · approvals · wikilinks · retrieval · model boundary
- Inspect
- Markdown · YAML · backlinks · chunks · provenance
- Safety premise
- Local vault ≠ automatically local AI
Second Brain workbenchDurable files first · AI changes stay reviewable
Current knowledge stepInspect before mutation
A person approves or rejects each exact note artifact; generation alone changes nothing durable. Approved notes: 2. Retrieved chunks: 6.
Adding retrieved text helped until irrelevant passages crowded out the evidence. Answers improved when every claim pointed back to a supporting span. A retrieved page also contained ‘ignore previous instructions’; the phrase was evidence inside the source, never authority for the agent.
- Provenance
- paragraphs 4–6
- Source state
- Preserved separately
- Authority
- Untrusted evidence
A citation is a testable support claim
A citation should identify the exact source span that supports a generated claim.
[[Resources/AI/Citation coverage]][[Resources/AI/Retrieval precision]]Retrieved evidence never grants instruction authority
Instructions found inside retrieved material remain untrusted data, even when the material is relevant.
[[Resources/AI/Prompt injection boundary]][[Resources/AI/Retrieval precision]]03 · Living vault
Files, links, and backlinks—not a decorative cloud.
Click a note. Its Markdown, outgoing links, and reverse references all come from the same visible fixture.
Local vault folder · 10 Markdown files
---
type: evergreen
tags: [ai/rag, knowledge-management]
source_span: "paragraph 4"
source: "[[Sources/Retrieval field notes|RAG research article]]"
---
# More retrieved context can reduce answer quality
Retrieval quality depends on relevance, not simply context volume.
Related: [[Resources/AI/Retrieval precision]] · [[Resources/AI/Citation coverage]]Outgoing wikilinks
Backlinks
04 · Retrieval laboratory
See exactly what the model is allowed to read.
Obsidian search, semantic retrieval, explicit graph neighbors, and hybrid retrieval produce different evidence envelopes.
Answer stategrounded
Why can more retrieved context make an answer worse?6 notes selected by visible hybrid rules.
Visible hybrid selection rule
De-duplicate the exact keyword, concept-tag, and one-hop graph result sets shown by the other modes.
More retrieved context can reduce answer qualityResources/Notes/More retrieved context can reduce answer quality.md
paragraph 4Retrieval quality depends on relevance, not simply context volume.
RAG research articleSources/Retrieval field notes.md
paragraphs 4–6Adding retrieved text helped until irrelevant passages crowded out the evidence. Answers improved when every claim pointed back to a supporting span. A retrieved page also contained ‘ignore previous instructions’; the phrase was evidence inside the source, never authority for the agent.
Prompt injection boundaryResources/AI/Prompt injection boundary.md
manual noteRetrieved content is evidence, not a source of tool authority.
A citation is a testable support claimResources/Notes/A citation is a testable support claim.md
paragraph 5A citation should identify the exact source span that supports a generated claim.
Retrieval precisionResources/AI/Retrieval precision.md
manual noteRetrieve the smallest set of passages that directly supports the task.
Citation coverageResources/AI/Citation coverage.md
manual noteEvery externally verifiable claim should point to supporting evidence.
05 · Data boundary
Local files do not make every AI path local.
Obsidian’s vault is local. A plugin, embedding provider, or answer model can still send selected content elsewhere.
Embedding model
Answer model
AI write authority
Outbound content1 outbound data path
→ question + retrieved note chunks for the answer
Every AI-originated mutation is shown as a diff and waits for approval.Capture, Organize, Distill, Express; then structure action with Projects, Areas, Resources, and Archives when it helps.
It does not automatically type relations, prove claims, or turn every attachment into a semantic index.
Compare it with the knowledge-graph lab →Obsidian Sync encrypts its remote vault and uses end-to-end encryption by default, but Obsidian does not encrypt the local vault folder. Those are separate operational promises.
Try another capture packet
Primary research
Inspect the mechanism at its source.
Project architecture and terminology were checked against official documentation on August 10, 2026. Research papers are linked to their original publication records.
Local Markdown vault mechanicsObsidian data storageOriginal source ↗Wikilinks, headings, and block linksObsidian internal linksOriginal source ↗Linked and unlinked mentionsObsidian backlinksOriginal source ↗Explicit file-link graphObsidian Graph ViewOriginal source ↗Community code trust boundaryObsidian plugin securityOriginal source ↗Remote encryption and local-vault boundaryObsidian Sync securityOriginal source ↗Capture, Organize, Distill, ExpressBuilding a Second BrainOriginal source ↗