ctxkit-ai 1.3.2

Context portability and knowledge ingestion tool — bridge your markdown vault to any subscription LLM

ctxkit — Context Portability and Knowledge Ingestion Tool

Version: 1.3.2 | Status: Implementation

A local, LLM-agnostic context delivery and knowledge sync tool for users who live across multiple subscription LLMs.

---

What It Does

ctxkit sits between your personal markdown knowledge base and whichever subscription LLM you are using at any given moment (Claude, ChatGPT, Gemini, or any other). It does two things:

Flow A — Context Retrieval: Given a topic, it retrieves relevant chunks from your vault, synthesises them into a coherent, cited briefing (using a local LLM), and writes the result to <vault>/ctxkit-output/. You open the file, copy it, and paste it as the opening context of your LLM conversation. Zero API calls to your subscription LLM. Zero cost beyond your existing subscription.

Flow B — Knowledge Ingestion: Given any document — a session summary from an LLM conversation, a research note, an article, a strategy doc, or any new piece of knowledge — it classifies the content against your existing vault, decides whether to update an existing note or create a new one, routes it to the right folder, detects conflicts, generates a proposal, and — only after your explicit approval — writes the changes to your vault and reindexes automatically.

---

Documentation

• Getting Started — installation, first index, first search, first ingest, daily workflow
• Config Reference — every config field documented with defaults, valid values, and tuning guidance
• Architecture — component design and data flow
• Retrieval & Eval Guide — symptom-to-fix guide for tuning retrieval quality

---

Installation

macOS — Homebrew

brew tap piyush-tyagi-13/ctxkit
brew install ctxkit

Any platform — shell script (installs uv + ctxkit)

curl -fsSL https://raw.githubusercontent.com/piyush-tyagi-13/context-portability-tool/master/install/install.sh | bash

Manual (if you already have uv or pipx)

uv tool install ctxkit-ai       # preferred
# or
pipx install ctxkit-ai

Ollama models (for local inference)

ollama pull nomic-embed-text   # embeddings
ollama pull qwen3.5:4b         # primary LLM — classification + proposals
ollama pull phi4-mini          # synthesis — fast, non-thinking

After install

ctxkit init       # interactive setup wizard
ctxkit deps install   # install any backend packages not yet present
ctxkit index      # index your vault

---

Commands

ctxkit init                           # Interactive setup wizard — create config
ctxkit index                          # Scan vault, show diff, confirm, index delta
ctxkit search <topic>                 # Synthesise briefing → write to <vault>/ctxkit-output/ (Flow A)
ctxkit search <topic> --raw           # Raw excerpts only — skip synthesis
ctxkit search <topic> --verbose       # Show chunk scores alongside results
ctxkit ingest                         # Paste any document — classify, route, propose (Flow B)
ctxkit ingest --file doc.md           # Ingest from a file (session summary, article, notes, etc.)
ctxkit map                            # Generate vault folder map for doc routing
ctxkit status                         # Show index health and drift warnings
ctxkit eval [topic]                   # Run quality evaluation checklist
ctxkit config                         # Open config file in editor
ctxkit config --validate              # Validate config and report errors

Multiple config profiles

ctxkit search "istio auth" --config ~/.ctxkit/config-technical.yaml
ctxkit search "career goals" --config ~/.ctxkit/config-personal.yaml

---

Quick Start

# 1. Configure (interactive wizard)
ctxkit init
# → asks for vault path, owner name, LLM backend, models
# → detects Ollama + pulled models, gives hardware-appropriate suggestions
# → writes ~/.ctxkit/config.yaml

# 2. Index your vault
ctxkit index

# 3. Retrieve context for an LLM conversation
ctxkit search "Bruno ingress path adaptor"
# → writes <vault>/ctxkit-output/2026-04-25-bruno-ingress-path-adaptor.md
# → open file, copy contents → paste into Claude/ChatGPT/Gemini

# 4. Ingest any document into your vault
ctxkit ingest --file my-session-summary.md   # LLM session summary
ctxkit ingest --file oss-strategy.md         # standalone research doc
ctxkit ingest                                # paste content directly
# → ctxkit classifies, routes to right folder, proposes changes → approve

---

Architecture

YOUR MARKDOWN VAULT
        │
        ▼
   ctxkit core
   ┌──────────┐  ┌──────────┐  ┌────────────┐
   │ Indexer  │  │Retriever │  │  Ingester  │
   └──────────┘  └──────────┘  └────────────┘
   ┌──────────┐  ┌──────────┐  ┌────────────┐
   │  Writer  │  │LLM Layer │  │VectorStore │
   └──────────┘  └──────────┘  └────────────┘
        │
   (copy-paste by user)
        │
        ▼
ANY SUBSCRIPTION LLM (Claude · ChatGPT · Gemini · Others)

ctxkit never talks to your subscription LLM directly. It prepares context (Flow A) and processes output from it (Flow B). The user is the bridge.

LLM calls in Flow A (retrieval): one local call to synthesise_model (default: phi4-mini via Ollama) to reformat raw excerpts into a coherent briefing. No calls to your subscription LLM. Skip with --raw to make Flow A fully LLM-free.

LLM calls in Flow B (ingestion): only when classification is ambiguous (score between low and high thresholds). Clear-match updates and clear new-file cases require no LLM call.

---

Configuration Reference

See config.yaml.example for the full annotated config. Key sections:

Section	Key fields	Purpose
vault	path, owner_name	Vault root path, owner identity for multi-person vaults
indexer	chunk_size, heading_levels	Chunking strategy and quality filters
embeddings	backend, local_model	Local (Ollama) or API-backed embeddings
retriever	top_k, similarity_threshold	Candidate retrieval, assembly, signposting
ingester	similarity_threshold_high/low	Classification thresholds, conflict detection
writer	append_position, backup	Append position, frontmatter injection, backups
llm	model, synthesise_model	Primary LLM (classify/propose) + synthesis model (search)
cli	theme, verbose	Terminal UI behaviour

---

Hardware Tiers

Hardware	LLM Model	Embedding Model
Apple M2 Air 16GB	qwen3.5:4b	nomic-embed-text
i5 + RTX 4070	qwen3:8b	bge-m3
Low-end / no GPU	gpt-4o-mini / claude-haiku-4-5	text-embedding-3-small

---

Project Structure

ctxkit/
├── cli/commands.py              # Typer commands, Rich rendering
├── core/
│   ├── indexer/                 # VaultScanner, ManifestManager, TextSplitter, ...
│   ├── retriever/               # KeywordPreFilter, VectorSearcher, ChunkStitcher, ...
│   ├── ingester/                # ClassificationEngine, ConflictDetector, ...
│   └── writer/                  # BackupManager, FrontmatterInjector, FileWriter, ...
├── llm/llm_layer.py             # classify() and propose() — single LLM abstraction
├── store/vector_store.py        # ChromaDB wrapper — 4 operations
├── config/                      # Pydantic models + YAML loader
└── utils/                       # Logging, file utilities

---

What ctxkit Is Not

• Not a chatbot or RAG question-answering agent
• Not an API wrapper around subscription LLMs
• Not a note-taking application
• Not an always-on background service
• Never writes anything without your explicit approval

---

ctxkit - Context Portability and Knowledge Ingestion Tool v1.3.2