MicroResolve

A pre-LLM reflex layer for AI agents — intent routing and tool prefilter in tens of microseconds. Embedded library with optional HTTP server.

v0.1 — early release. API may change before 1.0. Pin exact versions in production.

Documentation · Benchmarks & methodology · Changelog · Contributing

"the dispute on order 1834 came back invalid — refund the customer,
 mark the dispute resolved, then post in #ops that we lost it"

  → create_refund        (confirmed, high)
  → update_dispute       (confirmed, high)
  → slack_send_message   (confirmed, medium)
  → relation: sequential

                          ~92 µs · narrowed 129 tools to top-3
                          for the LLM, with no LLM call yet

Your LLM is System 2 — slow, deliberate, expensive. MicroResolve is System 1 — fast, learned, cheap. Pattern-match the easy 80% of queries in microseconds; let the LLM reason about the hard 20%. MicroResolve learns continuously from LLM-judged corrections — every misroute the LLM fixes flows back into the index, so the candidate set sharpens from production traffic alone.

Use it for:

• LLM-agent tool prefiltering — narrow 100+ tools to the top 3 the LLM actually needs to see this turn. Cuts prompt tokens, cuts latency, scales to large catalogs.
• Customer-support triage — route incoming tickets / chat messages to the right queue or workflow before the LLM gets involved (or without an LLM at all for the easy cases).
• Intent classification — single-utterance bucketing for conversational interfaces, IVR-style menus, search-query routing.
• Slash / chat command routing — recognize the user's command from free-form phrasing without retraining a model every time the catalog changes.
• Workflow / decision routing — multi-intent decomposition with relation detection (sequential / conditional / parallel / negation) for steps that need to fan out or chain.
• Permission and risk gating — classify a request into a risk tier before paying for an LLM round-trip.

[!IMPORTANT] MicroResolve is a prefilter, not an absolute classifier. It returns ranked candidates — the LLM is the final picker. Your agent prompt must include a confirm_full_catalog fallback tool so the LLM can reach the full catalog when none of the candidates fit. That fallback is part of the design, not an optional safety net. See Confirm-turn pattern.

Benchmarks

Agent tool routing — 129 real tools imported from 5 production MCP servers (Stripe / Linear / Notion / Slack / Shopify), single namespace, scored as a prefilter for an LLM picker:

+ auto-learn = incremental Hebbian + LLM-judged phrase ingestion from production corrections. No retraining pipeline; the data updates in place. Reproduce: python3 benchmarks/agent_tools_bench.py (~$0.55 / ~10 min on Haiku 4.5).

Methodology, datasets, and reproduction scripts live in benchmarks/. What's not yet benchmarked: out-of-scope rejection (see Confirm-turn pattern), adversarial robustness, drift over multi-week production traffic.

Single-utterance intent classification (academic baselines):

Quick start

Embedded (Rust):

use microresolve::Engine;

let engine = Engine::open("./data")?;
let ns = engine.namespace("agent")?;
ns.add_intent("cancel_subscription", &["cancel my plan", "stop the recurring billing"])?;
let r = ns.resolve("end my subscription right now");
// r.ranked[0].id == "cancel_subscription"

HTTP server (with the optional Studio UI):

cp .env.example .env                        # set LLM_API_KEY
cargo run --release --bin server --features server -- --data ./data
cd ui && npm install && npm run dev         # http://localhost:3000

Bootstrap intents from an existing spec:

curl -X POST http://localhost:3001/api/import/mcp/apply \
  -H "Content-Type: application/json" \
  -H "X-Namespace-ID: agent" \
  -d '{"tools_json": "<MCP tools/list response>", "selected": ["..."], "domain": ""}'

Bindings: microresolve on crates.io, microresolve on PyPI, microresolve on npm.

How it works

A query passes through five layers, all in-process, all in-memory:

raw query
  → L0  typo correction        (character n-gram Jaccard)
  → L1  vocabulary bridging    (morphology / abbreviation, OOV-gated)
  → L2  intent scoring         (IDF-weighted sparse term graph)
  → L3  confidence + disposition
  → L4  cross-provider tiebreak
  → result

failures → review queue → auto-learn worker → L1 + L2 updated

L0 fixes typos against the known vocabulary. L1 substitutes morphological variants (canceling → cancel) and abbreviations (pr → pull request) only when the source word is out-of-vocabulary, so distinctive user input is never rewritten. L2 is the core scorer: each intent has a learned sparse vector of IDF-weighted terms; multi-intent uses a token-consumption pass to confirm one intent then re-score the rest. L3 classifies the score distribution as confident / low_confidence / escalate. L4 breaks ties when the same action exists across multiple providers.

Every confirmed routing reinforces term weights. Every correction shifts weights away from the wrong intent. No retraining; the data updates in place.

Confirm-turn pattern (System 1 → System 2)

MicroResolve returns ranked candidates; the LLM picks one or falls back to the full catalog when nothing fits. Always include a confirm_full_catalog fallback tool in your agent prompt — without it, out-of-scope queries and novel phrasing route to the wrong tool:

Candidate tools (from the prefilter):
  - tool_a  (...description...)
  - tool_b  (...description...)
  - tool_c  (...description...)

If one of these clearly applies, call it. Otherwise — unsure, out of
scope, or candidates look unrelated — call `confirm_full_catalog` to
receive every tool, then pick.

The prefilter shrinks 150 tools to 3 in ~60 µs; the LLM is the final picker. Every confirmed call flows back as a corrected example, so the candidate set sharpens over time.

HTTP API

Send X-Namespace-ID: my-namespace to isolate intents per namespace. The core endpoints are /api/route_multi, /api/intents, /api/training/{review,apply}, and /api/import/mcp/{search,fetch,apply}. Full reference in the server API docs.

Multi-intent, projection, and multilingual

Native multi-intent — a single query can confirm several intents in one call, with the relation between them detected:

"cancel my order and update my address"
  → cancel_order   (confirmed, high)
  → update_address (confirmed, medium)
  → relation: parallel

Detected relations: sequential, conditional, negation, parallel.

Projected context — co-occurrence tracking discovers what auxiliary intents typically fire alongside the primary one, so your orchestrator can pre-fetch them in parallel without an extra LLM round-trip:

"I want a refund"
  → refund_order  (confirmed, high)
  → projected:    check_balance (21 %), warranty_lookup (13 %)

These relationships emerge from accumulated routing — they're not configured.

Multilingual — Latin scripts via whitespace tokenization, CJK (Chinese / Japanese / Korean) via Aho-Corasick automaton + character bigrams, all in the same namespace. Per-language seed phrases per intent. Multi-intent decomposition runs after tokenization, so a Chinese query like "取消订阅然后退款" decomposes the same way an English query does.

Import formats

• MCP tools/list (and Smithery registry passthrough)
• OpenAPI / Swagger — each endpoint becomes an intent
• OpenAI function-calling schemas
• LangChain tools

License

Dual-licensed under MIT or Apache-2.0 at your option — the standard Rust ecosystem licensing. Both are fully permissive and allow commercial use.

• LICENSE-MIT
• LICENSE-APACHE — adds an explicit patent grant

Contribution

Unless you state otherwise, any contribution intentionally submitted for inclusion in this work shall be dual-licensed as above, without any additional terms or conditions.

Commercial support

Maintained by Gladius Thayalarajan — consulting on custom integrations, multilingual tuning, and agent-stack tool routing: gladius.thayalarajan@gmail.com.