Semlith Files

The store learns the shape of the code in it, and search uses it.

Added

• A code graph, extracted on the same pass that re-embeds a file. Symbols and the defines, calls, imports, references and contains edges between them, read out of the syntax tree by tree-sitter, for Rust, TypeScript, Python, Go, Java and C. It hangs off the blake3 hash-change path that already drives re-embedding, so an edit under semlith start updates the graph in the same pass that updates the vectors. There is no build step and no artifact that can quietly go stale — which is the failure every snapshot-based code graph has, and the reason this one is not a snapshot.
• Extracted or inferred, on every edge. A call whose name the file also imports was resolved by the file itself; a bare name match was not. Two functions called new in different modules is the normal case in real code, not a corner case, so the difference is recorded and shown everywhere an edge appears. Nothing presents the second as the first.
• semlith symbol, semlith neighbors, semlith path, semlith impact, and the same four as the MCP tools semlith_symbol, semlith_neighbors, semlith_path and semlith_impact. Where a symbol is defined; what calls it and what it calls; the shortest chain of edges between two symbols; and what breaks if this one changes. impact is the answer to the most expensive mistake an agent makes on a real codebase — patching the call site that was reported and leaving every sibling caller broken.
• Graph-aware search. The top vector and keyword hits are mapped to the symbols in them, expanded one hop, and the chunks those neighbours live in become a third ranked list fused at the same weight as the other two. It costs no embedding and no model call, and it reaches the case neither existing list can: a concept spread across files that share no vocabulary. Every hit now says which lists found it — v, f, g on the command line — so a result the graph alone reached is legible as a neighbour of a match rather than a match.
• A Graph page and an Impact page in the portal. The graph on a canvas with a force layout that can be paused, panned, zoomed and dragged, beside a rail carrying the selected symbol's callers, callees and chunks. Reverse reachability as a table and as rings by hop, with a path finder. Both are free on every tier, now and after 0.13.0 makes the product paid.
• An opt-in retrieval ledger. semlith start --ledger records every query an agent ran into a hash-chained retrievals table inside the store: the query, the client, the hits, the excerpt tokens they actually read and the whole-file tokens a grep loop would have cost. Each row carries the hash of the row before it, so an edited or deleted row is detectable rather than merely unlikely. semlith ledger --last N prints it, the Ledger page shows the totals and the measured ratio, and none of it needs a key or leaves the machine. It is off unless asked for: a local tool that starts recording what you searched for without being told to is not meaningfully different from one that phones home.

Fixed

• The install script's progress bar. curl -L --progress-bar draws a bar for every hop of a redirect, and a hop carries no content length, so curl fell back to its bouncing #=O=- spinner. A GitHub release URL always redirects, so every install showed that before the real bar arrived, and a working download looked like line noise. The final URL is resolved first, which gives one request against a known size and one bar from 0% to 100%.

Notes

• Existing stores open unchanged. format_version stays 2 and nothing migrates: the new tables are created by the same IF NOT EXISTS batch every open already runs, so a store written by 0.11.0 opens with an empty graph and fills it on the next index pass, and a store written by 0.12.0 still opens under 0.11.0. Both directions are tested against the released 0.11.0 binary. docs/compatibility.md explains why the number did not move.
• Measured. Extraction costs 98 ms over 589 KiB of source producing 757 symbols and 5,536 edges — 0.17% of the 58 s that corpus takes to index, because embedding dominates and tree-sitter is fast. The six grammars add 4.6 MB to the binary, 38.6 MB to 43.2 MB.
• Six languages carry edges. Everything else is searchable exactly as before, with no symbols. The About page says which is which.
• Everything in this release is free on every tier, permanently. 0.13.0 introduces paid tiers by adding views above these, never by locking one.

Install

macOS and Linux

:::sh
curl -fsSL https://raw.githubusercontent.com/semlith/semlith/main/install.sh | sh

Windows

:::powershell
irm https://raw.githubusercontent.com/semlith/semlith/main/install.ps1 | iex

Or download the archive for your platform below, unpack it, and put semlith on your PATH. Intel macOS has no prebuilt binary: ONNX Runtime no longer publishes x86_64 macOS builds.

Verify a download against SHA256SUMS before running it.