Releases·Tool Release
tool-prune prunes 50+ tool schemas to candidates in 0.4ms, cutting prompt tokens 92% with zero dependencies
A developer released tool-prune, a dependency-free JS/Python library that prunes large tool-schema lists before the model call. It targets local 7B/8B models, where 50+ schemas degrade attention and cause distractor hallucinations, and avoids the extra LLM round-trip that in-band tool_search requires.
7 engagement1 sourceFri, Sep 18, 2026, 04:55 PM
AI Overview
- tool-prune filters tool schemas client-side pre-flight via router.filter(), so 50+ schemas no longer degrade local 7B/8B model attention or cause distractor hallucinations
- Pruning runs offline in 0.4ms via FWHT in pure JS and Python (23µs with SIMD), with zero external dependencies
- Adds no LLM round-trips, unlike in-band tool_search, which costs an extra generation turn (~2s) and can trap small models in discovery loops
- Cuts prompt tokens by 92% and offers direct dispatch for deterministic queries in under 1ms
- Code and playground available on GitHub (hemanth/tool-prune) and hemanth.github.io/tool-prune
Entities
Related
Releases·Tool Release
Official LangChain Middleware: Give Your Agent 100 Tools. Expose 4. Keep the Same Accuracy.
A developer released langchain-dynamic-tools-middleware, an open-source project now part of the LangChain ecosystem, that uses local hybrid retrieval to show agents only the most relevant tools instead of every tool schema. On a 100-tool benchmark it cut prompt tokens by 80.5% with no accuracy loss on BFCL and 122× faster tool routing than LangChain's LLMToolSelectorMiddleware, with no added routing cost.
A developer released langchain-dynamic-tools-middleware, an open-source project now part of the LangChain ecosystem, that uses local hybrid retrieval to show agents only the most relevant tools instead of every tool schema. On a 100-tool benchmark it cut prompt tokens by 80.5% with no accuracy loss on BFCL and 122× faster tool routing than LangChain's LLMToolSelectorMiddleware, with no added routing cost.
1 engagement1 sourceredditTue, Sep 15, 2026, 12:39 PM
Social·Community
Homelab agent drops vector DB for closed-vocab token-overlap tool selection — recall ties dense, cost drops
A developer running a self-hosted assistant (local Gemma planner, 96 tools) removed the BGE-M3 embedding and vector search from tool selection, replacing it with token overlap over compositional tool names plus four typed CPU rules. Recall matched the dense baseline (0.786 vs ~0.76 R@5) at ~10 ms with no model, and prefiltering to 8 tools cut frontier-model input tokens by ~92%. A public reproducible benchmark backs the numbers, with caveats that the approach depends on owning a closed tool vocabulary.
A developer running a self-hosted assistant (local Gemma planner, 96 tools) removed the BGE-M3 embedding and vector search from tool selection, replacing it with token overlap over compositional tool names plus four typed CPU rules. Recall matched the dense baseline (0.786 vs ~0.76 R@5) at ~10 ms with no model, and prefiltering to 8 tools cut frontier-model input tokens by ~92%. A public reproducible benchmark backs the numbers, with caveats that the approach depends on owning a closed tool vocabulary.
18 engagement1 sourceredditThu, Sep 17, 2026, 10:01 PM
Developments·Paper
AgentWeave: Routing Before Reasoning for Efficient Function Calling in Tool-Rich Language Models
AgentWeave is a deterministic pre-inference routing layer that reduces the candidate action space before language-model inference, improving efficiency in tool-rich function-calling scenarios. It constructs a bounded model-visible action space using eligibility and requirement checks, leaving the downstream model unchanged.
AgentWeave is a deterministic pre-inference routing layer that reduces the candidate action space before language-model inference, improving efficiency in tool-rich function-calling scenarios. It constructs a bounded model-visible action space using eligibility and requirement checks, leaving the downstream model unchanged.
0 engagement1 sourcearxivMon, Aug 24, 2026, 10:38 AM
Social·Community
Why are we sending every tool schema on every agent turn??
A developer reports their agent sends forty tool schemas on every turn when a router could exclude thirty-eight before the model sees a token, inflating costs. They are testing dynamic tool routing and schema pruning but face tradeoffs between recall and token cost, and ask how others decide which tools enter each prompt turn.
A developer reports their agent sends forty tool schemas on every turn when a router could exclude thirty-eight before the model sees a token, inflating costs. They are testing dynamic tool routing and schema pruning but face tradeoffs between recall and token cost, and ask how others decide which tools enter each prompt turn.
88 engagement2 sourcesredditThu, Sep 10, 2026, 08:05 PM
Products·concept
OntoPrune: neuro-symbolic context pruning for faster local LLM inference
OntoPrune is an open-source library that prunes LLM context using neuro-symbolic techniques before prompting, cutting tokens by 83-86% without causing hallucinations. It targets prompt-evaluation bottlenecks when running small coding models via Ollama on CPU-only hardware. It helps local-LLM users on laptops or mini PCs get much faster time-to-first-token.
OntoPrune is an open-source library that prunes LLM context using neuro-symbolic techniques before prompting, cutting tokens by 83-86% without causing hallucinations. It targets prompt-evaluation bottlenecks when running small coding models via Ollama on CPU-only hardware. It helps local-LLM users on laptops or mini PCs get much faster time-to-first-token.
1 engagement1 sourceredditMon, Oct 5, 2026, 12:02 PM