Personal Research Vault

## Inspiration

I had 340 PDFs on my NAS — research papers, lease agreements, product
manuals, saved articles — and no way to search across them. Not really
search. Ctrl+F only works if you already know which file to open.

Cloud tools exist for this. NotebookLM, Perplexity, ChatGPT with file
upload. But my lease, my tax documents, my medical records — I'm not
uploading those anywhere. The whole point of a NAS is that my data stays
mine.

So the question became: why does the intelligence have to live in the cloud
when the data already lives locally?

## What It Does

Personal Research Vault turns any NAS folder into a private, cited
knowledge base. You ingest a folder once. Then you ask questions — forever.

Example:

"What does my lease say about subletting?"

"Subletting requires written landlord approval at least 30 days in
advance and is subject to a 10% rent increase for the sublet period.
[Source: lease-2024.pdf, ~p.3]"

Every answer is grounded exclusively in your own documents. The Skill
physically cannot cite something that isn't in your vault — a citation
guard strips any hallucinated references before the response reaches you.

Nothing leaves the NAS. No API keys. No cloud calls.

## How It's Built

The Skill chains five native UGOS capabilities across two modes:

INGEST pipeline:
File Agent (discover files) → Document Agent (extract text) → Indexing
Agent (chunk at ~400 words with 10% overlap, store in Knowledge Agent
under a scoped vault:{name}:* namespace). A manifest tracks every
ingested file by name and modification time, making re-ingestion
incremental — only changed files are reprocessed.

QUERY pipeline:
Dual retrieval — Knowledge Agent semantic search + Universal Search
secondary pass — surfaces the most relevant chunks. These go into the
Uliya LLM with a strict grounding prompt. A post-generation citation
validator checks every cited filename against the retrieved context and
strips anything that wasn't actually there.

## Why This Only Makes Sense on a NAS

Three things make this NAS-native, not just NAS-compatible:

1. Privacy by architecture. Sensitive documents never leave the device.
   The local LLM means even inference is on-device.
   
2. Persistence. The knowledge base lives in the Knowledge Agent store
   permanently. Unlike cloud tools that forget between sessions, your vault
   grows over time.
   
3. Lifetime data accumulation. People's most important documents —
   contracts, records, research — already live on their NAS. The vault
   meets the data where it is.
   

## Challenges

Incremental ingestion without duplicates. Re-running INGEST naively
would re-index everything. The solution: store a manifest in the Knowledge Agent itself, keyed by filename + mtime. Only files that are new or
modified get reprocessed.

Keeping the LLM honest. Retrieval-augmented generation still hallucinates
citations. The citation guard cross-references every source the LLM names
against the actual retrieved chunks — if a filename wasn't in the context
window, it gets stripped and flagged to the user.

Graceful failure on bad files. Encrypted PDFs and image-only scans
return empty text. The Skill catches these at the Document Agent step, logs
them with reasons, and continues — a bad file should never abort an entire
1,000-document ingest.

## What's Next

OCR support for image-only scanned PDFs via AI Album — the one gap in the
current version. And cross-vault queries for users who want to search
across multiple topic-scoped vaults in a single question.

---

Built With

• agent
• ai
• document
• file
• knowledge
• llm
• markdown
• nas
• rag
• search
• ugos
• uliya
• universal