context-engineer-tabtii 0.1.0

Context-Engineer for local AI Agents. Zero-config RAG for Ollama with automatic chunking, token-budget awareness, and source citation.

ConText

Context-Engineer for local AI Agents. Zero-config RAG for Ollama with automatic chunking, token-budget awareness, and source citation.

pip install context-engineer
ollama pull nomic-embed-text
ollama pull llama3.2

context build ./docs
context query "What is the main topic?"

Why ConText?

Most RAG pipelines force you to choose between LangChain's complexity, ChromaDB's infrastructure, and naive vector search's poor results. ConText is the third option: opinionated, local-first, zero-config.

Feature	ConText	LangChain	LlamaIndex	Naive RAG
Setup time	30 seconds	30 minutes	20 minutes	5 minutes
Auto chunking	✅ 5 strategies	⚠️ DIY	✅	❌ Fixed only
Token-budget aware	✅	⚠️ DIY	⚠️ DIY	❌
Source citation	✅ Built-in	⚠️ DIY	⚠️ DIY	❌
Runs offline	✅ Ollama only	✅	✅	❌
Dependencies	5 packages	30+	20+	varies
Storage	SQLite	optional	optional	in-memory

Features

• 🎯 5 Chunking Strategies — Auto-selected per file type (markdown, code, semantic, fixed, sliding)
• 💰 Token-Budget Manager — Fits retrieved chunks to your model's context window
• 📚 Source Citation — Every answer includes [1], [2], [3] references
• 🚀 Zero Dependencies — Just pip install context-engineer + Ollama running locally
• 💾 SQLite Storage — No external vector database needed
• 🔌 HTTP API — MCP-compatible server for IDE integration
• 📊 Built-in Benchmarks — Compare against naive RAG baseline

Quick Start

1. Install

pip install context-engineer

2. Install Ollama models

# Embedding model (required)
ollama pull nomic-embed-text

# LLM model (pick one)
ollama pull llama3.2          # 2GB, fast
ollama pull qwen2.5-coder    # 4GB, code-specialized
ollama pull mistral          # 4GB, general

3. Index your docs

Option A: Local files

context build ./docs

This recursively finds all .md, .txt, .py, .js, .ts, .go, .rs, .java, .c, .cpp, .h, .hpp, .json, .yaml, .yml files and indexes them with the right chunker.

Option B: Crawl a website

context crawl https://docs.example.com --max-pages 20

Respects robots.txt, BFS-crawls same-domain pages, removes script/style/nav/footer.

4. Ask questions

context query "What guarantees memory safety in Rust?"

Output:

╭─ Answer (27295ms, 714+97 tokens) ─────────────────╮
│ Rust's ownership system guarantees memory safety   │
│ at compile time. Every value has an owner, and     │
│ there can only be one owner at a time. When the    │
│ owner goes out of scope, the value is dropped [1]. │
╰────────────────────────────────────────────────────╯
Sources
┏━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━┓
┃ # ┃ Source               ┃ Strategy ┃ Score ┃
┡━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━┩
│ 1 │ rust-intro.md        │ markdown │ 0.855 │
│ 2 │ rust-intro.md        │ markdown │ 0.765 │
└───┴──────────────────────┴──────────┴───────┘

Python API

from context_engineer import ConText

# Initialize
ctx = ConText(
    db_path=".context/store.db",
    llm_model="llama3.2",
    embed_model="nomic-embed-text",
    context_window=8192,
)

# Index files
ctx.build("./docs")

# Add raw text
ctx.add_text("Some important info", source="notes.md")

# Query
answer = ctx.query("What is X?")
print(answer.text)        # "X is ... [1]"
print(answer.sources)     # [Source(...), ...]
print(answer.latency_ms)  # 1234

# Stats
ctx.stats()
# {'documents': 5, 'chunks': 27, 'embeddings': 27, ...}

HTTP API

context serve --port 8765

# Query
curl -X POST http://localhost:8765/query \
  -H "Content-Type: application/json" \
  -d '{"question": "What is X?"}'

# Build
curl -X POST http://localhost:8765/build \
  -H "Content-Type: application/json" \
  -d '{"path": "./docs"}'

# Stats
curl http://localhost:8765/stats

Architecture

ConText Pipeline
    │
    ├─ Document Loader (.md, .py, .js, ...)
    │       │
    │       ▼
    ├─ Auto-Selected Chunker
    │       │
    │       ├─ Markdown   (headers, code blocks)
    │       ├─ Code       (functions, classes)
    │       ├─ Semantic   (sentence boundaries)
    │       ├─ Fixed      (N chars, M overlap)
    │       └─ Sliding    (larger overlap)
    │       │
    │       ▼
    ├─ Embedder (Ollama nomic-embed-text + LRU cache)
    │       │
    │       ▼
    ├─ SQLite Store (chunks + vectors as BLOB)
    │
    └─ Query
        │
        ├─ Cosine-Similarity (top-20)
        │
        ├─ Token-Budget Greedy-Fit
        │
        └─ LLM Generation (Ollama llama3.2)
            │
            └─ Cited Response [1][2][3]

CLI Reference

context build <path>          Index documents
context query "question"      Ask a question (with sources)
context stats                  Show database stats
context list                   List indexed documents
context remove <source>        Remove a document
context serve                  Start HTTP API server

Global options:

• --db PATH — SQLite database location (default: .context/store.db)
• --llm MODEL — Ollama LLM model (default: llama3.2)
• --embed MODEL — Ollama embed model (default: nomic-embed-text)
• --ollama-url URL — Ollama API (default: http://localhost:11434)
• --context-window N — LLM context window (default: 8192)

Roadmap

• v0.1.0 — Core RAG + 5 chunkers + CLI + HTTP API
• v0.2.0 — Hybrid search (BM25 + vector)
• v0.3.0 — Cross-encoder reranking
• v0.4.0 — Multi-modal (image + text)
• v0.5.0 — Cloud sync + team features (Pro)
• v1.0.0 — Stable API

License

MIT © 2026 Torben

See Also

• Ollama — Local LLM runtime
• Architecture docs
• Status & Roadmap
• Benchmark results