/token-reducer
Build a compact context packet with preprocess, BM25-ranked FTS retrieval, adaptive vector fallback, reranked top-K chunks (default 3 from settings), relevance-floor filtering, and compression.
> /plugin marketplace add Madhan230205/token-reducer > /plugin install claude-token-reducer@Madhan230205-claude-token-reducer
How it fires
How this command gets triggered: by you, by Claude, or both.
- Fires itselfClaude auto-loads it when your prompt matches the work.
- You can call itInvoke it directly when you want it.
- Slash command
/token-reducer
Context preview
What this command does when you run it.
Build a compact context packet with preprocess, BM25-ranked FTS retrieval, adaptive vector fallback, reranked top-K chunks (default 3 from settings), relevance-floor filtering, and compression.
Command definition
token-reducer.mddescription: Build a compact context packet with preprocess, BM25-ranked FTS retrieval, adaptive vector fallback, reranked top-K chunks (default 3 from settings), relevance-floor filtering, and compression.
argument-hint: query [--inputs path1,path2] [--top-k 3-5] [--relevance-floor 0.15-0.25] [--hybrid-mode fallback|always] [--session-id id] [--json]
allowed-tools: [Read, Glob, Grep, Bash, Task]
token-reducer
Run the full token-slim pipeline and return a context packet suitable for Claude reasoning.
Input
User arguments:
`$ARGUMENTS`
Execution plan
1. If no explicit input paths are provided, default to current workspace root. 2. Run pipeline script in `run` mode. 3. Enforce FTS-first retrieval with BM25 scoring. 4. Run vector retrieval adaptively (fallback default), merge results, and keep top **K** reranked chunks (plugin default **3** via `settings.json` → `defaultTopK`). 5. Apply **`relevanceFloor`** so low-scoring chunks are not summarized. 6. Compress before response handoff.
Primary command
`python "${CLAUDE_PLUGIN_ROOT}/scripts/context_pipeline.py" run --inputs . --query "$ARGUMENTS" --hybrid-mode fallback --session-id default --top-k 3`
For stricter cost control, keep `--top-k 3` and a **specific** query (e.g. “Find the JWT validation function”) so irrelevant files score out before compression.
Notes
- Prefer local pipeline to stay free-cost.
- ANN acceleration is used automatically when `hnswlib` + `numpy` are available.
- Use Context7 only when external library docs are needed.
- Always return citations and estimated savings.
- Large raw pasted query blobs are guarded; put logs/files in `--inputs`.
- Default chunk size, word budget, `top-k`, and relevance floor are loaded from the plugin **`settings.json`** when `CLAUDE_PLUGIN_ROOT` is set (repo `settings.json` when running from a checkout).
Read more
description: Build a compact context packet with preprocess, BM25-ranked FTS retrieval, adaptive vector fallback, reranked top-K chunks (default 3 from settings), relevance-floor filtering, and compression. argument-hint: query [--inputs path1,path2] [--top-k 3-5] [--relevance-floor 0.15-0.25] [--hybrid-mode fallback|always] [--session-id id] [--json] allowed-tools: [Read, Glob, Grep, Bash, Task]
token-reducer
Run the full token-slim pipeline and return a context packet suitable for Claude reasoning.
Input
User arguments:
`$ARGUMENTS`
Execution plan
1. If no explicit input paths are provided, default to current workspace root. 2. Run pipeline script in `run` mode. 3. Enforce FTS-first retrieval with BM25 scoring. 4. Run vector retrieval adaptively (fallback default), merge results, and keep top **K** reranked chunks (plugin default **3** via `settings.json` → `defaultTopK`). 5. Apply **`relevanceFloor`** so low-scoring chunks are not summarized. 6. Compress before response handoff.
Primary command
`python "${CLAUDE_PLUGIN_ROOT}/scripts/context_pipeline.py" run --inputs . --query "$ARGUMENTS" --hybrid-mode fallback --session-id default --top-k 3`
For stricter cost control, keep `--top-k 3` and a **specific** query (e.g. “Find the JWT validation function”) so irrelevant files score out before compression.
Notes
- Prefer local pipeline to stay free-cost.
- ANN acceleration is used automatically when `hnswlib` + `numpy` are available.
- Use Context7 only when external library docs are needed.
- Always return citations and estimated savings.
- Large raw pasted query blobs are guarded; put logs/files in `--inputs`.
- Default chunk size, word budget, `top-k`, and relevance floor are loaded from the plugin **`settings.json`** when `CLAUDE_PLUGIN_ROOT` is set (repo `settings.json` when running from a checkout).
⚡ Cut Claude token usage by 90%+ — free, open-source, local-first context compression for Claude Code. Hybrid RAG (BM25 + ONNX vectors), AST chunking, reranking. No API needed.
Repo: Madhan230205/token-reducer