NervaPack

NervaPack

[PyPI version] [Python Versions] [License: MIT] [Token Reduction] [Verified] [MCP Registry]

NervaPack is a privacy-first, offline knowledge graph for your codebase. It solves two fundamental problems with standard Vector RAG:

• Token waste — chunk-based RAG retrieves blobs of text that may only tangentially relate to your query, bloating your context window.
• Privacy risk — sending code to cloud embedding APIs leaks your proprietary logic.

NervaPack runs 100% on your machine. It uses tree-sitter to parse your codebase into a deterministic Abstract Syntax Tree graph, then uses a local Ollama model to draw hard semantic edges between your documentation and your code. Queries traverse this graph with a K-Hop BFS, returning a hyper-targeted, token-efficient context window — no cloud required.

Verified Performance

91.2% Average Token Reduction — independently verified on real-world codebases.

Cost Savings: $181–$724 per developer per year (GPT-4o to Claude Sonnet)

📊 View Full Benchmarks · 🧪 Messy Code Performance

Code quality impact: 90–99% reduction on clean code, 50–75% on legacy/messy code. Even poorly structured codebases benefit significantly.

Why NervaPack vs. standard Vector RAG

Prerequisites

• Python 3.10+
• Git — your project must be a git repository (git init if not)

(Optional for semantic code-doc binding) — an LLM provider. Structural graph indexing and basic queries work out-of-the-box with zero configuration and no cloud connection.

Installation

# Recommendedpip install nervapack
# With optional featurespip install "nervapack[mcp]"          # MCP server for Claude Code / Cursorpip install "nervapack[memory]"       # agent memory MCP serverpip install "nervapack[metrics]"      # exact token counts (tiktoken)pip install "nervapack[dashboard]"    # web dashboard (streamlit + plotly)pip install "nervapack[claude]"       # Claude API supportpip install "nervapack[openai]"       # OpenAI API supportpip install "nervapack[all]"          # everything

On first run, ChromaDB downloads an ONNX embedding model (~30 MB) to ~/.cache/chroma/. This is a one-time download.

Quick Start

cd your-project/
# 1. Build the knowledge graph (runs locally, no LLM required for basic use)nervapack ingest .
# 2. Query for context — get focused results + token savings dashboardnervapack query "How does authentication work?"
# 3. Add semantic doc-to-code edges (requires an LLM)nervapack enrich .
# 4. Visualize the full graphnervapack visualize --enhanced --communities
# 5. After changing files, sync incrementally (fast — only changed files)nervapack sync .
# 6. If you ingested wrong data or need a fresh startnervapack clean --allnervapack ingest .
# 7. Check system healthnervapack doctor

Command Reference

nervapack ingest [PATH] — Build the graph

Scans PATH (default: .) and builds the full knowledge graph.

What happens:

1. Walks the directory tree with tree-sitter, skipping dist/, build/, node_modules/, venv/, site/, .tox/, and dozens of other build directories automatically.
2. Parses source files into exact AST nodes: classes, functions, imports.
3. Chunks all .md files by header hierarchy.
4. Embeds every entity into a local ChromaDB vector store (ONNX by default, Ollama optional).
5. Optionally binds doc chunks to code nodes via an LLM, adding EXPLAINS edges.
6. Saves the graph once to .nervapack/graph.graphml.

Re-ingesting is safe — upsert is used throughout, so running ingest twice does not duplicate data.

nervapack ingest .                           # auto-detect LLMnervapack ingest . --llm ollama              # force Ollamanervapack ingest . --llm claude              # use Claude APInervapack ingest . --llm openai --model gpt-4o-mininervapack ingest . --embeddings ollama       # use Ollama for embeddings too

Supported languages (bundled): Python, JavaScript, JSX, TypeScript, TSX

Additional languages:

pip install "nervapack[go]"            # Gopip install "nervapack[rust]"          # Rustpip install "nervapack[java]"          # Javapip install "nervapack[c]"             # C / C headerspip install "nervapack[cpp]"           # C++pip install "nervapack[ruby]"          # Rubypip install "nervapack[csharp]"        # C#pip install "nervapack[all-languages]" # all of the above

Exclude directories — create .nervapackignore in your project root (gitignore syntax):

dist/build/site/generated/*.egg-info/

nervapack query PROMPT — Query the graph

Retrieves focused context for a natural-language prompt and prints a token savings dashboard.

What happens:

1. Intent detection — "what breaks if I change X" routes to impact analysis (reverse BFS); exact symbol names bypass vector search.
2. ChromaDB returns the top-3 most semantically similar nodes.
3. Those nodes seed a K-Hop BFS through the NetworkX graph (default: 1 hop, both directions).
4. Adjacent nodes — including any markdown docs via EXPLAINS edges and memory notes via TOUCHES edges — are collected.
5. A focused Markdown context block is printed, ready to paste into an LLM prompt.
6. Token efficiency panel shows savings vs. naive "dump the whole file" RAG.

nervapack query "How does authentication work?"nervapack query "What calls VectorStore?"nervapack query "what breaks if I change GraphBuilder"  # impact analysis

Example output:

Query: "How does authentication work?"
Query Router: Intent: semantic, Direction: bothVector Search: Found 3 seed nodes
Retrieved Context:──────────────────────────────────────────────────────────# NervaPack Context Retrieval## File: `src/auth/middleware.py`### CLASS: AuthMiddleware (L15-L42)...──────────────────────────────────────────────────────────
╭──────────────  NervaPack Token Efficiency  ──────────────╮│  Strategy              Tokens   Reduction                 ││  Naive RAG (3 files)   12,840   100% (base)              ││  NervaPack              1,180     9.2%                    ││ ─────────────────────────────────────────────────────────││  Tokens saved: 11,660   Reduction: 90.8%                 ││  Cost saved (GPT-4o $2.50/1M): $0.0292 per query         │╰───────────────────────────────────────────────────────────╯

nervapack sync [PATH] — Incremental update

Updates only the files that changed since the last ingest. Uses GitPython to diff the working tree.

nervapack sync .

A full ingest on a large project can take minutes. sync turns that into a 2–5 second surgical update per file. Re-parses changed files, batch-upserts new vectors, and saves the graph once at the end.

nervapack clean [OPTIONS] — Remove ingested data

Wipe graph data and start fresh. Use this when you have duplicate vectors, ingested the wrong directory, or need to reduce disk usage.

nervapack clean --vectors          # wipe ChromaDB only (chroma_db/)nervapack clean --graph            # delete graph.graphml onlynervapack clean --history          # clear query + graph history logsnervapack clean --all              # everything above (keeps memory.db)nervapack clean --all --yes        # skip confirmation (CI / scripts)

Never deleted by clean: memory.db — your agent memory is always safe.

Typical workflow after a bad ingest:

nervapack clean --allnervapack ingest .

nervapack enrich [PATH] — Add semantic edges

Runs LLM doc-to-code binding on an already-ingested graph. Use this if you:

• Ran ingest without an LLM and want to add EXPLAINS edges now.
• Added new documentation and want to bind it without a full re-ingest.

nervapack enrich .                           # auto-detect LLMnervapack enrich . --llm ollama --model llama3nervapack enrich . --llm claude              # shows cost estimate before proceeding

nervapack status [--detailed] — Graph health

nervapack status            # node/edge counts + unsynced filesnervapack status --detailed # full analytics: health score, language distribution, coverage

Health score (0–100) factors in documentation coverage, node connectivity, graph density, and edge diversity. A structural-only graph typically scores 30–40; after enrich it rises to 70–90.

nervapack visualize [OPTIONS] — Interactive HTML graph

nervapack visualize                           # basic visualizationnervapack visualize --enhanced                # + real-time search + path findernervapack visualize --enhanced --communities  # + community detection (Louvain)nervapack visualize --output ~/my-graph.html  # custom output pathnervapack visualize --no-browser              # generate without opening

Produces a standalone HTML file with no external dependencies — drag, zoom, search, find shortest paths between nodes.

nervapack explore TARGET [--hops N] — Focused subgraph

Extract and visualize the N-hop neighbourhood of a specific file, class, or function.

nervapack explore GraphBuilder --hops 2nervapack explore src/auth/middleware.py --hops 1nervapack explore "function:parse"

nervapack dependencies [FILE] — Import dependency analysis

Analyze file-level import chains, detect circular dependencies, and visualize the dependency graph.

nervapack dependencies                        # full project analysisnervapack dependencies src/graph/builder.py   # single filenervapack dependencies --no-cycles            # skip cycle detection

nervapack hotspots [OPTIONS] — Change frequency analysis

Show which files change most often in git history — prime targets for documentation and review.

nervapack hotspots                            # top 20 by commit countnervapack hotspots --since "6 months ago"    # recent history onlynervapack hotspots --ext .py --churn         # Python files, sort by lines changed

nervapack history [OPTIONS] — Query history

nervapack history              # last 10 queries with token savingsnervapack history --limit 50   # last 50nervapack history --stats      # aggregate: total savings, cost avoided, top topicsnervapack history --clear      # delete history

nervapack serve [--port N] — Web dashboard

nervapack serve                # http://localhost:8501nervapack serve --port 8080    # custom port

Requires nervapack[dashboard]. Shows graph overview, language distribution, analytics, query history trends, and an interactive graph explorer.

nervapack doctor — Environment check

nervapack doctor

Verifies Python version, tree-sitter grammars, embedding backend, Ollama connectivity, and MCP config. Run this after installation or when troubleshooting.

Storage Layout

Everything lives in .nervapack/ inside your project root:

.nervapack/├── graph.graphml          # NetworkX DiGraph (AST + EXPLAINS edges)├── chroma_db/             # ChromaDB vector store (ONNX embeddings)├── memory.db              # Agent memory (SQLite + FTS5, bi-temporal)├── query_history.jsonl    # Per-query token savings log└── graph_history.jsonl    # Ingest/sync event log

Add .nervapack/ to .gitignore to keep it out of version control.

Disk usage guide:

• chroma_db/ — typically 10–100 MB depending on project size. Run nervapack clean --vectors && nervapack ingest . if it grows unexpectedly.
• graph.graphml — typically 0.5–5 MB.
• memory.db — grows with agent usage; rarely exceeds a few MB.

MCP Integration (Claude Code, Cursor, Windsurf)

NervaPack ships two MCP servers in the same package. Drop this .mcp.json in your project root:

{  "mcpServers": {    "nervapack": {      "command": "nervapack-mcp",      "description": "NervaPack knowledge graph (v0.6.1) — query, graph_status, explore, impact"    },    "nervapack-memory": {      "command": "nervapack-memory-mcp",      "description": "NervaPack agent memory (v0.6.1) — store, recall, and reason over facts across sessions"    }  }}

Knowledge Graph MCP tools (nervapack-mcp)

Agent Memory MCP tools (nervapack-memory-mcp) — 17 tools

CLAUDE.md template

Add to your CLAUDE.md to wire NervaPack into every Claude Code session:

## Always use NervaPack MCP tools
At the start of every session:1. Call `memory_start_session("<task name>")` to open a named session.2. Call `memory_recall("project context", budget_tokens=400)` to load prior decisions.3. Call `query("<topic>")` before answering any question about how the code works.
During the session — call `memory_store` for:- Any decision made (kind="decision" with rationale and alternatives_rejected)- Any fact discovered about system behaviour (kind="fact")- Any procedure or convention established (kind="procedure")
At session end — call `memory_end_session("<one-paragraph summary>")`.

Architecture

nervapack ingest .       │       ├─ ASTParser (tree-sitter)               16 extensions, 10 languages       │    └─ ParsedEntity[]: class, function, import       │       ├─ GraphBuilder (NetworkX DiGraph)       │    ├─ Nodes: file, class, function, import, markdown       │    ├─ Edges: DEFINES (AST, confidence=1.0)       │    ├─ Edges: REFERENCES (heuristic name overlap, confidence=0.7)       │    └─ Edges: EXPLAINS (LLM or keyword binding, confidence=0.5–0.9)       │       ├─ VectorStore (ChromaDB)       │    ├─ Embedding: ONNX (default) or Ollama       │    └─ upsert — re-ingest is idempotent       │       └─ GraphHistory  — snapshot recorded after each ingest/sync
nervapack query "..."       │       ├─ Intent detection (impact / exact / semantic)       ├─ VectorStore.search() → seed node IDs       ├─ GraphRetriever.retrieve_context()       │    └─ deque BFS, direction: forward / reverse / both       ├─ MemoryStore.get_touches_for_file() → memory context injection       └─ TokenMeter → savings vs. naive RAG
nervapack clean --all       └─ Deletes chroma_db/, graph.graphml, history logs          Never touches memory.db

Key source modules:

nervapack.memory — Agent Memory

Stop re-pasting project context into every new chat.

# Store a decisionmemory_store(    "Chose JWT over session cookies for auth_service — stateless horizontal scaling",    kind="decision",    entities=["auth_service"],    confidence=0.9,    rationale="Stateless tokens enable horizontal scaling without shared session store.",    alternatives_rejected=["server-side sessions", "PASETO"],)
# Recall in any future sessionmemory_recall("project context", budget_tokens=400)

## Memory recall: "project context" (6 items · 171/400 tokens)
### Decisions- [d_0019f2] 2026-06-05 · conf 0.90 — Chose JWT for auth_service
### Facts- [f_0019f3] 2026-06-05 · conf 1.00 — auth_service issues 15-min access tokens
### Procedures- [p_0019f5] 2026-06-12 · conf 1.00 — Deploy: GitHub Actions → staging → prod manual

Token efficiency:

CLI:

nervapack-memory init                          # create schemanervapack-memory stats                         # counts + top entitiesnervapack-memory search "JWT auth"             # FTS searchnervapack-memory timeline "auth service"       # chronological tracenervapack-memory audit d_0019f2abc             # access audit trailnervapack-memory rebind old/path.py new/path.py  # update file links after renamenervapack-memory export --out dump.json        # JSON dump

Data model: 8 node kinds, 7 edge kinds, bi-temporal (valid_from/valid_until), never hard-deletes by default.

Privacy

NervaPack is 100% offline by default:

• Embeddings use ChromaDB's built-in ONNX model (runs on CPU, no cloud).
• LLM calls (when using Ollama) go to localhost:11434 only.
• All graph and vector data lives in .nervapack/ inside your project.
• No telemetry, no analytics, no network calls at runtime.

Only if you explicitly pass --llm claude or --llm openai does any code leave your machine.

Contributing

1. Fork the repo and create a branch.
2. Run tests: python3 -m pytest tests/memory/ -q
3. Run docs check: python3 -m mkdocs build --strict
4. Open a pull request against master.

Bug reports and feature requests: issue tracker.

License

MIT — see LICENSE.