resonance-lattice 0.10.0

Portable project memory cartridges for AI assistants

Resonance Lattice

Portable semantic knowledge infrastructure for AI assistants.

Resonance Lattice applies semantic-model thinking to unstructured knowledge, turning docs and code into portable .rlat knowledge cartridges you can search, inspect, compose, diff, and use across assistant workflows.

Build local knowledge cartridges for retrieval, memory management, skill integration, fine-grained context control, context priming, and offline or LLM-free use.

pip install resonance-lattice
rlat build ./docs ./src -o project.rlat
rlat search project.rlat "how does this project work?" --format text

---

What Is Resonance Lattice?

Resonance Lattice is a local semantic knowledge system for unstructured knowledge and AI assistants. It turns docs, code, and other file collections into portable .rlat knowledge cartridges.

A knowledge cartridge is a semantic model packaged as a file. You can search it, inspect it, compose it with other cartridges, diff versions, and use it across assistant workflows without standing up a hosted retrieval stack.

It is not a hosted retrieval service, not a vector database, and not an LLM. It is portable knowledge infrastructure: a queryable layer between your source material and the assistants or tools that use it.

For the fuller product and architecture story, see Overview, Cartridge Architecture, and Semantic Model.

---

Why Use Resonance Lattice

Use Resonance Lattice when you want retrieval that behaves like infrastructure: portable, inspectable, composable, and local. It sits between raw source material and the assistants or tools that need grounded access to it.

If you use	What it is good at	What Resonance Lattice adds
grep	Exact text match	Semantic retrieval, inspection, and reusable cartridges
Standard RAG	Hosted or index-backed retrieval	Portable file artifacts, local control, and composition
LLM direct	Reasoning and generation	Grounded evidence from your own sources
Note vaults	Human-authored organization	Automatic semantic modeling over the files you already have

1. Your knowledge is a file

A cartridge is a single .rlat artifact you can move, version, archive, diff, and share like any other file. The knowledge layer is portable, not tied to a service dependency.

2. It is composable for fine-grained context control

Cartridges are building blocks, not monoliths. Build per-domain cartridges for docs, code, design, or compliance and combine them per question. Composition happens locally at query time.

See Context Control for the deeper composition model, reusable context setups, and cartridge algebra workflows.

Operation	What it does	Guarantee
merge	Combine cartridges into a unified field	Commutative, associative
diff	Surface what changed between two snapshots	Directional signed delta
forget	Remove a knowledge subset cleanly	Algebraically precise
project	View one cartridge through the lens of another	Semantic projection, not keyword filtering
contradict	Find where two cartridges disagree	Symmetric divergence detection

# Multi-cartridge search
rlat search docs.rlat "auth flow" --with code.rlat

# View code through a compliance lens
rlat search code.rlat "data handling" --through compliance.rlat

# What changed since baseline?
rlat search current.rlat "what changed?" --diff baseline.rlat

3. It is inspectable

You do not just get top-k results. Resonance Lattice exposes structural diagnostics that standard vector-database workflows do not usually surface directly:

Command	What it tells you
rlat xray corpus.rlat	Corpus-level health: signal quality, saturation, diagnostic flags
rlat locate corpus.rlat "query"	Where a query sits in the knowledge landscape and what the field does not cover
rlat probe corpus.rlat <recipe>	Quick insight recipes for novelty, saturation, coverage gaps, and more

4. It is local-first

No hosted service, no API keys, and no LLM required. The first build downloads the default encoder from Hugging Face; after that, building and querying can run fully locally. Pair it with a local LLM such as Ollama or llama.cpp for a fully private stack.

5. It is mathematical, not generative

The field has algebraic structure. Adding a source, removing it, merging cartridges, and diffing versions are defined operations with stable behavior. Removing a file returns the field to the same state it had before that file was added. Merge is order-independent. The same inputs produce the same field. No temperature, no sampling, no drift.

---

Quick Start

Python >=3.11 is required. The first build downloads the default encoder. Install onnxruntime if you want faster local encoding.

pip install resonance-lattice
pip install onnxruntime  # optional

Build a cartridge from your docs and code:

rlat build ./docs ./src -o project.rlat

Search it:

rlat search project.rlat "how does auth work?" --format text

Inspect what the cartridge covers:

rlat profile project.rlat

When you want assistant-native setup, initialize the project and wire in the defaults:

rlat init-project --auto-integrate

This auto-detects common project inputs, builds .rlat/project.rlat, generates .claude/resonance-context.md, creates .rlat/manifest.json, and with --auto-integrate updates .mcp.json and injects a cartridge section into CLAUDE.md.

For the full walkthrough - including profiling, comparison, MCP integration, assistant context files, and HTTP serving - see Getting Started.

---

Assistant Integration

These paths are composable rather than exclusive. You might use MCP and CLI side by side, and layer cartridge-backed skills or instruction files on top when you want dynamic, targeted context for different applications of the same skill.

See CLI, MCP, API Reference, and Skills Integration and Architecture for the detailed interface docs.

MCP Server

Use the MCP server when your assistant supports MCP and you want cartridge search and diagnostics available as native tools inside the conversation.

Assistant	Config file
Claude Code	.mcp.json in project root
VS Code / GitHub Copilot	.vscode/mcp.json
Cursor	cursor/mcp.json or Settings > MCP

{
  "mcpServers": {
    "rlat": {
      "command": "rlat",
      "args": ["mcp", "project.rlat"]
    }
  }
}

Once loaded, the assistant can:

• search or generate LLM-ready context from a cartridge
• compose multiple cartridges, diff them, and inspect freshness or discovery state
• run diagnostics such as profile, compare, locate, and xray
• route into cartridge-backed skill workflows when those are configured

Because the server keeps the cartridge and encoder warm in memory, repeated queries avoid the cold-start path.

CLI

Use the CLI when the assistant has terminal access or when you want zero extra configuration. This works with Claude Code, GitHub Copilot, Cursor, Codex, Cline, Windsurf, and any other tool that can run shell commands.

rlat search project.rlat "how does auth work?" --format json
rlat resonate project.rlat "what are the design constraints?" --format context
rlat profile project.rlat

When you want LLM-ready output, --mode controls how the context is framed:

Mode	What it tells the LLM
augment	Use your own knowledge, but add detail and citations from these sources
constrain	Answer ONLY from the provided sources — if it's not covered, say so
knowledge	Base your answer primarily on this context; be transparent about gaps
custom	Your own system prompt, your rules

rlat resonate project.rlat "how does auth work?" --mode constrain --format context

Skills and Instruction Files

Use cartridge-backed skills and instruction files alongside MCP or CLI when you want the same skill to keep its workflow structure while its knowledge changes with the request. This is the key advantage of the skills system: dynamic, targeted context instead of the same static background on every run.

A skill becomes cartridge-backed by adding a few frontmatter fields:

---
name: fabric-notebook-ingest
description: Create Fabric notebooks for data ingestion...
cartridges:
  - .rlat/fabric-docs.rlat
  - .rlat/pyspark-docs.rlat
cartridge-queries:
  - "How do you create a notebook in Fabric through the API"
  - "Fabric workspace authentication and authorization patterns"
cartridge-mode: augment
cartridge-budget: 2000
---

When a cartridge-backed skill triggers, context can load in four tiers:

Tier	Source	What it provides
1. Static	SKILL.md header	Workflow structure, templates, decision trees
2. Foundational	Skill-authored queries	Baseline knowledge the skill always needs (40% of budget)
3. Specific	User query	Context unique to this request (30% of budget)
4. Derived	Caller-supplied queries	Implicit needs the user didn't express (30% of budget)

Tiers 1-3 are automatic. Tier 4 accepts additional queries from the caller via --derived, which lets an orchestrating agent surface knowledge the user did not explicitly ask for. Skills without cartridge-* fields work exactly as before. See docs/SKILL_INTEGRATION.md for the full architecture.

You can also generate a supplemental context file and reference it from your instruction file:

rlat summary project.rlat -o .rlat/resonance-context.md

Tool	Instruction file	Reference syntax
Claude Code	CLAUDE.md	@.rlat/resonance-context.md
GitHub Copilot	.github/copilot-instructions.md	@.rlat/resonance-context.md
Cursor	.cursorrules	@.rlat/resonance-context.md

How It Complements Other Layers

Layer	Role
Instruction files	Rules and project conventions
Skills	Executable capabilities and tool-use patterns
Memory	Persistent facts across conversations; Resonance Lattice can also back this layer when you index notes, sessions, or other artifacts
Resonance Lattice	Queryable project knowledge and adaptive context

Resonance Lattice is complementary to the other layers. It gives them a reusable semantic knowledge layer they can query, inspect, and adapt to the task at hand.

---

Common Workflows

Search and inspect a cartridge

Use search when you want evidence-backed retrieval, and profile when you want to understand the semantic shape of the cartridge. search supports text, json, context, and prompt formats depending on whether you are reading, scripting, or feeding an assistant.

rlat search project.rlat "how does auth work?" --format text
rlat search project.rlat "how does auth work?" --format json
rlat profile project.rlat

Compare versions or knowledge domains

Use compare when you want overlap and coverage differences, and diff when you want a queryable semantic delta.

rlat compare baseline.rlat current.rlat
rlat diff current.rlat baseline.rlat -o delta.rlat

Update cartridges as sources change

Use add when you want to append new sources, and sync when you want the cartridge to track file additions, updates, and deletions across source directories.

rlat add project.rlat ./new_docs
rlat sync project.rlat ./docs ./src

Compose context across cartridges

Use composition when the answer depends on more than one knowledge domain.

rlat search docs.rlat "auth flow" --with code.rlat
rlat search code.rlat "data handling" --through compliance.rlat
rlat search current.rlat "what changed?" --diff baseline.rlat

Generate and share assistant context

Use summary when you want a compact assistant primer, and export --field-only when you want to share the semantic model without the embedded evidence text.

rlat summary project.rlat -o .rlat/resonance-context.md
rlat export project.rlat -o shared.rlat --field-only

---

How It Works

A knowledge cartridge is a semantic model packaged as a file. rlat build chunks your sources, encodes them into a shared semantic space, and writes the result to a .rlat cartridge that can be queried later.

Each cartridge contains three layers:

Layer	What it does
Field	Stores the semantic model of the corpus
Registry	Maps semantic hits back to ranked sources
Store	Returns evidence text, passages, and metadata

At query time, the question is encoded into the same space. The field and registry identify the best matches, and the store returns the evidence needed for reading, scripting, or assistant injection.

The encoder is part of the contract. The cartridge records the encoder fingerprint used at build time, and query commands normally restore it automatically.

You can package cartridges in two ways:

Store mode	What you get	Best for
embedded (default)	Self-contained .rlat with field, registry, and evidence text	Portability, sharing, and archiveability
external	.rlat with field and registry only; evidence is read from local source files via --source-root	Smaller artifacts or workflows where source text should stay outside the cartridge

For deeper inspection, Resonance Lattice also exposes xray, locate, probe, and RQL. Those are advanced tools for diagnostics and programmable field operations, not required for normal use.

---

Benchmarks

Resonance Lattice is not "a new embedding model." Its value comes from the portable cartridge, the retrieval pipeline, and the assistant workflow.

• On a 24,635-chunk Microsoft Fabric documentation corpus with 100 evaluation questions, the full rlat retrieval pipeline reached Recall@5 1.00, MRR 0.93, and 0% failed retrieval.
• On a 2,266-file corpus, rlat search returned 24.6x fewer tokens than a grep + read top 5 files workflow while keeping ranked passages and source attribution intact.
• In LLM-grounding evaluation, feeding rlat context reduced hallucinations from 78% to 16% and raised fact recall from 0.27 to 0.91.
• On five BEIR datasets, the full pipeline beat flat E5 on three datasets and came within 97% on SciFact. Results vary by corpus, which is why fit to your own data matters more than any single leaderboard.

For full methodology and extended results, see Benchmarks, Benchmark Runbook, and benchmarks/results/.

---

Practical Limits

The .rlat cartridge stays stable as the user-facing abstraction, but a few implementation limits are worth knowing.

What	Detail	What to do
Initial build is CPU-intensive	First build encodes every chunk through E5-large-v2	Incremental sync only re-processes changed files. ONNX runtime (pip install onnxruntime) provides 2-5x CPU speedup. CUDA GPU is supported if available.
Default encoder is English-optimised	Non-English text can retrieve less reliably with the default setup	The backbone is configurable with --encoder, but multilingual alternatives should be revalidated on your corpus.
Best benchmark numbers use the full pipeline	The strongest retrieval numbers come from lexical injection plus reranking on factual and technical corpora	Full pipeline adds latency, but rlat search defaults to auto reranking and stays on the dense path when results are already well-separated.

---

Status

Alpha (0.9.0)

See Status and Boundaries for shipped surfaces, experimental areas, and current limits.

Docs

• Overview
• Getting Started
• CLI Reference
• MCP
• Context Control
• Encoder Guide
• RQL Reference
• Benchmarks
• FAQ

License

Apache 2.0.