Silica Core

Lightweight, local evidence retrieval tools for code and research

Silica locates the source, symbol, page, or passage you and your agents need. Maximum signal, minimum machinery.

Silica indexes the markdown, code, PDFs and office files under one root and serves them to Claude Code, Cursor, Hermes, Codex, OpenCode and every other popular harness, as MCP tools or as shell commands that print the same JSON. A hit is a path, a section, a line, a window of text and the numbers to judge it by. The harness owns the loop.

Three commands, then a question asked the way you would ask any other. The harness calls silica_search, and the answer carries the file, the page and the line it came from.

Install

In Claude Code, install the plugin instead: it brings its own server, dense leg included.

uv tool install 'silica-core[mcp]'         # BM25 over documents and their sections (faster, lighter)

uv tool install 'silica-core[mcp,dense]' # in addition the dense leg: numpy, a static model (still fast, more precise)

The second line adds the dense leg: numpy and a static model2vec model, no torch and no GPU. It stays inert until the model is named and the sections are embedded — the last stanza of the Quickstart. Take it when the questions are paraphrases that share no words with the text; an exact term or an identifier is answered by the lexical leg either way, and only that leg reports terms_absent.

pipx works the same. The package and the command are silica-core, the tools are silica_*.

Upgrading from 0.9 or older: the silica command is silica-core since 0.10.0 (silica belongs to silica-harness). A client registered by an older silica setup still launches silica mcp, which stops starting once uvx fetches 0.10.0. Run silica-core setup once and it rewrites that entry.

Quickstart

In any folder of markdown, code, PDFs or office files:

silica-core init                              # adopt the folder: ignore file, vault.yaml, first index; the MCP server serves adopted folders only

silica-core search "leveled compaction" -k 5 # the best located passages silica-core setup claude # register the MCP server, write the guidance block into ~/.claude/CLAUDE.md

optional, with the [dense] extra: the dense leg, a static model, nothing leaves the machine

export SILICA_EMBEDDING_MODEL=model2vec/minishlab/potion-retrieval-32M silica-core index --embed

Nine arXiv papers, indexed in 2.1 s. The hit names the file, the page and the passage; the page beside it is the check.

Tools

In a source tree every function, method, class and constant is its own unit: a hit's section is the symbol, span its lines, and silica_read(path, section=…) serves the body. For a symbol whose name is known, grep wins; for a question that names none, the search comes first: the tool description and the block silica-core setup claude writes say so. The contract, the reply shapes and the acceptance checks are in TOOLS.md. Nothing needs an API key or a network.

How search says no

A ranked list always has a top, even when the corpus does not answer. Three fields say how much the result is worth:

• coverage: the share of the query's idf mass the hit's matched terms carry. Near 1, every rare term matched; near 0, only common words did.
• terms_absent: query terms that occur nowhere in the corpus.
• matched_terms: the words this hit actually contains.

raft consensus log replication over the same nine papers: raft and consensus occur in none of them, coverage falls to 0.19, and the top hit is about data replication. On 254 papers the top hit of an answered question carries 0.69 to 1.00; a question the corpus does not cover, 0.44. Silica exposes the signals; the harness decides whether to stop, read or rephrase.

What a reply also says (0.12)

Asked in November 2026, nine models named what they need from a knowledge layer; each of these was kept only where a probe on a production model showed the failure it fixes (TOOLS.md has the numbers):

• scope counts every file the search could not read, an office file with no extracted .md beside it included, so a folder of .docx no longer reads as silent.
• A hit from a converted note carries source_state: stale when its PDF was replaced after the conversion, before the passage is quoted.
• terms_near offers the corpus's closest spelling of an absent word (quitanllu → quitanlu) instead of leaving a misspelt name to read as absent.
• unseen counts the ranked documents as strong as the weakest shown that were not shown; exclude takes the paths already read and serves the next ones, so a long list is paged instead of re-fetched.
• silica_files(since=…) lists only what changed after a time, dated: what changed since the last session.

Benchmarks

nDCG@10 on BEIR SciFact · NFCorpus, the same documents and queries for every arm. BEIR's published BM25 baselines are 0.665 · 0.325. Silica's lexical index needs no model; the others serve lexical search from an index that also holds embeddings.

Code, on the twenty SWE-QA questions zvec-grep publishes for its own benchmark, same embedder, k = 10, scored on the files and symbols the reference answer rests on.

On BEIR the two hybrids tie at the 95% interval: the same vectors rank the same, with no daemon and no vector store. On code, Silica's hybrid file MRR is +0.135 over zvec-grep's hybrid (95% interval +0.01 to +0.27), paired per question. The fusion also gains +0.21 MRR and +0.30 symbol hit over Silica's lexical arm; no reranker or graph expansion is involved.

On this measured scope, Silica is a compact, local, SOTA-competitive retriever: it matches zvec-grep on BEIR and leads the paired SWE-QA code-localization replay with the same embedder.

Retrieval matters only if the agent does less work without losing the answer. These are separate experiments and are not pooled:

The repository result is one repetition: turns improve by 1.75 (95% interval 0.55 to 3.05 fewer), while Judge and cost remain inconclusive. The document rows belong to a 144-run study over twelve questions and a 5.5M-token corpus. They establish less work on that workload, not a universal agent claim.

Corpora, intervals, per-task exceptions and reproduction commands are in benchmarks.

Harnesses

silica-core setup writes the registration into the client's own config and backs up what was there; for claude it also puts a guidance block, when to search before grep, into ~/.claude/CLAUDE.md. silica-core setup --list names the clients: claude, codex, cursor, windsurf, zed, cline, roo, continue, goose, opencode, openhands, gemini, antigravity, dsh, hermes, openclaw, agent-zero, claude-desktop, lmstudio, anythingllm and librechat; shell, python and generic print recipes for anything else. The server serves the folder the client opens in; --vault DIR or SILICA_VAULT fixes the root.

Every written block, and the Claude Code plugin, run silica-core mcp --retrieval local-hybrid: potion-retrieval-32M in the server process, index and vectors built in the background at start, the search lexical and dense: warming until they land. Nothing leaves the machine; the one download is the model, once. npx skills add kiycoh/silica-core installs the skills (the one that tells an agent when to reach for the tools, and five that explain, summarize, compare, schematize and diagram from the indexed documents), and nothing else. Shell recipes, Docker and t