Skip to content

TODOS.md

Latest commit

 

History

History
60 lines (47 loc) · 2.62 KB

TODOS.md

File metadata and controls

60 lines (47 loc) · 2.62 KB

Deferred work

Known limitations and possible future work. Not commitments — items here ship if and when they become worth the cost.

Search quality and features

Cross-encoder performance optimization

The MLX reranker (Qwen3-Reranker-0.6B) is still the main warm-query latency cost on realistic vault chunks. Passage text is now capped before reranking; remaining possible reductions:

  • Hardware-specific batching or sequence-classification reranker probes. Gate on real blind-test latency because MLX batch throughput depends on prompt length and padding.
  • Better query routing so only ambiguous queries pay the full rerank budget.

Additional CLI subcommands

Helpers exist inside seeklink/app.py but are not exposed on the CLI:

  • seeklink suggest-links <path> — find notes that should be linked.
  • seeklink resolve-suggestion <id> approve|reject — accept / reject a suggestion, writing [[link]] into the source on approval.
  • seeklink graph <path> --depth N — show the link neighborhood of a note.

Embedder upgrade path

Current default: jinaai/jina-embeddings-v2-base-zh (~330 MB, 768-dim). Stronger multilingual alternatives (Qwen3-Embedding-0.6B, BGE-M3, jina-v3) are >2 GB in ONNX form and not currently fastembed-supported. Revisit when a <500 MB multilingual embedder with better CJK scores becomes available in fastembed. SEEKLINK_EMBEDDER_MODEL is already configurable; a swap requires a full re-index.

Ingesting non-markdown sources

A vault sources/ folder could hold raw external content (PDFs, papers, textbooks) for semantic search. Requires a PDF→markdown pipeline (markitdown / marker / docling), optional LLM metadata extraction, and probably chapter-level chunking instead of the current fixed-size chunks.

Daemon and platform

Daemon freshness integration

Cold-start seeklink search / seeklink status emit stderr warnings when indexed files have drifted on disk. The daemon path does not propagate those warnings back to clients. Adding a warnings field to the daemon JSON response would let cli_client surface them in the same shape.

Multi-vault daemon support

The daemon binds to a single socket (~/.rhizome/seeklink.sock) regardless of vault. For multiple vaults to run concurrent daemons, hash the vault path into the socket name. Deferred until multi-vault is a real user need.

Linux reranker

The MLX reranker is Apple Silicon only. On Linux it self-disables. Options to restore reranking on Linux:

  • Port the scoring loop to onnxruntime (CUDA or CPU).
  • Run a GGUF build of the same model via llama.cpp (e.g. Mungert/Qwen3-Reranker-0.6B-GGUF).