FENRIS

An answer is useful.Its evidence is essential.

Search & retrieval

Retrieval & product engineering

In development · Local-first

Art-directed interface study · sample content

Fenris answers questions across a Google Drive library using a local search index. Each answer links back to the documents it used, so you can check the source yourself.

The problem

Searching a Drive library often means remembering the right filename before you can find the right fact. A fluent generated answer does not solve that problem unless its source is recoverable and its evidence is current.

The approach

I separated synchronization, retrieval, and answer generation. Read-only Drive access maintains a rebuildable local projection. Hybrid search finds evidence, verification checks it, and the answer layer exposes an explicit state when the evidence is missing, stale, or conflicting.

Your files, locally searchable.

The dashboard exposes source, chunk, and sync state. Google Drive remains the authoritative copy; the local index is a projection that can be rebuilt.

Follow the answer to its source.

The conversation carries citations with file titles, pages, and quoted evidence. A source identifier must resolve to retrieved text before it can appear as a citation.

Make the gaps visible.

The source ledger records indexed, skipped, and failed files. Supported parsing tiers are explicit, and the answer state distinguishes insufficient evidence from a system error.

How it’s built.

Commit the cursor with the indexed files.

The change-feed cursor and index writes commit together. A crash can replay work instead of advancing beyond files that never reached the index.

Combine keyword and vector search.

BM25 and vector candidates are fused, reranked, and filtered for diversity. Adaptive result counts follow the evidence rather than a fixed number of chunks.

A non-answer is a valid result.

The state ladder distinguishes ANSWERABLE, INSUFFICIENT, STALE, CONFLICTING, and ERROR. Retrieved documents are treated as evidence, never as instructions to execute.

Sync

Read-only Drive change feed

Retrieve

BM25 + vectors + reranking

Verify

Freshness and evidence state

Answer

Resolved citations or abstention

Built with
PythonReactTypeScriptSQLiteFTS5sqlite-vecDoclingGoogle Drive API

Interface walkthrough · sample data · no audio

Open the source. Check the answer.

A local retrieval workflow that pairs answers with inspectable source evidence and makes missing context visible.

Generation can use a configured provider or local model; local retrieval does not imply every operation is offline. PDF and image parsing depend on the installed tier.

Next project ↗

GETFINALOFFER

Recruiting intelligence