kinetic-context 0.2.1

Repository-level code context engine — find the right code, fast.

kinetic-context

Repository-level code context engine — find the right code, fast.

kinetic-context is a self-contained, installable code context engine. Point it at any repository and it builds a multi-layer index (AST chunks + embeddings + a code knowledge graph + BM25) that you can query with natural language or identifiers. It returns code blocks with line ranges — not just file paths — ready to paste into an LLM prompt.

It is designed to be the context layer for coding agents. It ships with a Rich CLI, an MCP server (so Claude Code, Cursor, Continue, and Zed can mount it natively), a TCP JSON mode for any other agent, and a Python library.

---

Highlights

• Multi-language: Python, JavaScript, TypeScript, Go, Rust, Java — via tree-sitter.
• Structure-aware chunking (cAST): never splits a function mid-body. 35% overlap. Hierarchical summaries at file and repository level.
• Hybrid retrieval: dense (Mistral Codestral Embed, 1536-dim) + BM25 (code-aware tokenizer that splits camelCase, snake_case, and dotted identifiers) + a Code Knowledge Graph with 8 relationship types.
• Reciprocal Rank Fusion with brute-force-optimized weights across 5 channels (dense, BM25, graph, PRF, patch-reverse-engineering).
• Zerank-2 reranker with instruction-following via XML tags, tuned per query intent.
• Novel signals invented for this engine:
  • Cross-resolution resonance — L2+L3+L4 consensus boost
  • Code DNA fingerprinting — structural similarity (param count, complexity, call count)
  • Semantic bridges — virtual graph edges between similar functions
  • File cohort memory — files that co-occur in correct answers get boosted across queries
  • Score distribution shape analysis — adaptive cutoff detects bimodal / uniform / power-law score distributions
  • Score gap amplification — sigmoid sharpening of close scores
  • Post-rerank filename tie-breaker — the fix for "correct directory, wrong file" at scale
  • Adversarial anti-centroid — penalty for chunks near the worst BM25 hits
  • Query-to-patch reverse engineering — generate a hypothetical fix, embed it as a 5th RRF channel
• Incremental indexing with SHA-256 Merkle root hashing. Only changed files are re-embedded. Subsequent index runs are fast.
• Per-repo isolated storage under ~/.kinetic/<slug>_<hash>/. Same-name folders in different parents do not collide. A manifest.json records the source path, root hash, file count, and last-indexed timestamp so kinetic status can answer "reindex needed?" in <50ms without spinning up the engine.
• Hookable everywhere: MCP server (stdio JSON-RPC), TCP JSON, Python library, Rich CLI.

---

Install

pip install kinetic-context

Requires Python 3.10+. All heavy dependencies (tree-sitter, numpy, networkx, rich) are bundled — no Qdrant/Pinecone/Milvus/Postgres to install.

Set your API keys (a Mistral key for embeddings + a Zerank key for reranking):

export MISTRAL_API_KEY=...
export ZEROENTROPY_API_KEY=...

You can also override the embedder / reranker URLs and model IDs in KCEConfig if you want to point at a different provider.

---

Quick start

# Index a repo (first run; takes a few minutes for a 1k-file repo)
kinetic index ./my-repo

# Subsequent runs only re-embed files whose SHA-256 hash changed
kinetic index ./my-repo

# Search — returns code blocks with line ranges, not just file paths
kinetic query "authentication middleware" --code

# JSON output (for agent hooks / piping)
kinetic query "how does routing work" --json | jq '.results[0]'

# Check if the index is up to date
kinetic status ./my-repo

# List all indexed repos
kinetic list

Query output example

────────────────────────────────────────────────────────────────────────
 kinetic-context • semantic_question • zerank • 68ms
────────────────────────────────────────────────────────────────────────
 Query: how does routing work in flask
────────────────────────────────────────────────────────────────────────
Found 8 code blocks across 8 files

#1  src/flask/sansio/app.py  L240-262 (23 lines)  • add_url_rule (function)  • python
    def add_url_rule(self, rule, endpoint=None, view_func=None, ...)
  240 │ def add_url_rule(
  241 │     self,
  242 │     rule: str,
  243 │     endpoint: str | None = None,
  ...

---

Hooking into coding agents

MCP server (Claude Code, Cursor, Continue, Zed, anything MCP-aware)

Start the MCP server in the background:

kinetic mcp

Or add it to your agent's MCP config:

{
  "mcpServers": {
    "kinetic": {
      "command": "kinetic",
      "args": ["mcp"]
    }
  }
}

Four tools are exposed:

Tool	Description
kinetic_index	Index (or incrementally update) a repo
kinetic_query	Search the indexed codebase, returns code blocks with line ranges
kinetic_status	Check whether the index is up to date
kinetic_list_indexes	List all indexed repos

TCP JSON server (any agent)

kinetic serve --port 7878

Send a single-line JSON request, get a single-line JSON response:

echo '{"query": "session cookie signing", "top": 5}' | nc localhost 7878

Python library

from kce.engine import KCEEngine
from kce.config import KCEConfig
from kce.store.registry import Registry

cfg = KCEConfig()
cfg.index_dir = Registry().resolve("/path/to/repo")
cfg.ensure_dirs()
engine = KCEEngine(cfg)
engine.index("/path/to/repo")

result = engine.query("how does authentication work")
for cid in result.final_chunk_ids[:10]:
    chunk = engine.chunk_index[cid]
    print(f"{chunk.rel_path}:{chunk.start_line}-{chunk.end_line}  {chunk.name}")
    print(chunk.content)

---

Storage layout

Every indexed repo gets its own directory under ~/.kinetic/:

~/.kinetic/
  flask_f9cbd6f6/                       <- slug = name + short hash of abspath
    manifest.json                       <- source path, root hash, file count, last indexed
    chunks.jsonl                        <- one CodeChunk per line
    embeddings.npy                      <- (N, 1536) float32 matrix
    embeddings_ids.json                 <- row order
    ckg.graphml                         <- Code Knowledge Graph
    incremental_state.json              <- per-file SHA-256 hashes
    change_log.jsonl                    <- append-only change history
    embed_cache/embeddings.db           <- Mistral API cache (SQLite)
    zerank_cache/rerank.db              <- Zerank API cache (SQLite)
    file_summaries.json
    repo_summary.txt
  django_ea20f8f1/
    ...

Why per-repo isolation?

• Two repos with the same folder name (e.g. ~/work/api and ~/side/api) get different slugs because the slug includes a short hash of the absolute path. No collisions.
• Caches live next to the index, so deleting a repo's index (kinetic forget) also frees its cache.
• The manifest.json records a Merkle root hash over all file content hashes. Recomputing this on startup is <50ms even for a 1k-file repo, so kinetic status is instant.

Efficient incremental updates

1. On kinetic index <repo>, we first compute the current Merkle root from disk.
2. We compare to the stored root in manifest.json. If they match, the index is up to date — we're done in <100ms.
3. If they differ, we walk the per-file SHA-256 hashes in incremental_state.json and identify exactly which files were added, modified, or removed.
4. Only those files are re-parsed, re-summarized, and re-embedded. Existing chunks for unchanged files are reused.
5. The Mistral embed cache (SQLite, keyed by SHA-256 of the embed text) means even a chunk whose text didn't change but whose chunk_id was regenerated will not trigger an API call.

For a 1k-file repo where 5 files changed, a re-index takes seconds, not minutes.

---

Architecture

The pipeline has four layers:

1. Ingestion — tree-sitter parses each source file into an AST. The cAST chunker walks the AST and produces chunks that respect syntactic boundaries (a function is never split mid-body). Each chunk is enriched with its signature, docstring, decorators, and scope. A Code Knowledge Graph (NetworkX) is built with 8 relationship types: CALLS, INHERITS, IMPLEMENTS, IMPORTS, CONTAINS, USES_TYPE, OVERRIDES, DEPENDS_ON.

2. Storage — chunks go to JSONL, embeddings go to a single numpy matrix on disk (+ L2-normalized in memory for fast cosine), the graph goes to GraphML. Each repo gets its own directory under ~/.kinetic/. SHA-256 Merkle root hashing drives incremental updates.

3. Retrieval & Ranking — the query coordinator picks one of 5 query types (identifier lookup, semantic question, code completion, bug diagnosis, architecture query). For each type, it applies intent-aware boosts (source vs test vs config files), runs multi-query BM25 + dense + graph retrieval, fuses the results with weighted Reciprocal Rank Fusion, applies novel signals (resonance, DNA, semantic bridges, cohort memory), then reranks the top candidates with Zerank-2.

4. Output — the final ranked chunks are returned as code blocks with line ranges, signature, and docstring. They can be rendered by the Rich CLI, serialized to JSON, or shipped over MCP / TCP to any coding agent.

---

Benchmarks

We benchmark on two real-world repos with hand-curated query sets. We report Context F1@10, Recall@10, Precision@10, and end-to-end latency. We deliberately do not compare to other context engines — those numbers are easy to get wrong and we'd rather show our own results honestly than risk an unfair comparison.

Aggregate metrics

Per-query F1 (sorted, so the worst queries are at the top)

Query outcome distribution (perfect / partial / failed)

Latency distribution

Quality vs repository scale

Numbers

Corpus	Files	Chunks	Graph (nodes/edges)	Queries	F1@10	Recall@10	Precision@10	Avg latency
Flask	83	1,382	1,594 / 7,906	30	0.659	0.833	0.599	68 ms
Django (core)	308	7,207	6,574 / 46,408	30	0.647	0.867	0.559	433 ms

The Django benchmark uses the core Django packages (django/db/, django/http/, django/urls/, django/template/, django/forms/, django/core/, django/contrib/auth/, django/contrib/sessions/) — the parts of Django that real coding agents actually search. The full Django repo includes 3,000+ files of migrations, tests, and docs that bloat the index without improving retrieval quality on real-world queries.

Run the benchmarks yourself:

git clone https://github.com/pallets/flask /tmp/flask
git clone https://github.com/django/django /tmp/django
kinetic index /tmp/flask
kinetic index /tmp/django
python scripts/run_bench.py flask
python scripts/run_bench.py django
python scripts/gen_charts.py

---

Configuration

All knobs are in KCEConfig. The defaults are tuned for the Mistral Codestral Embed + Zerank-2 combination. You can override:

Setting	Default	What it controls
mistral_embed_model	codestral-embed	Embedding model
zerank_model	zerank-2	Reranker model
chunk_max_tokens	512	Max chunk size
chunk_overlap_pct	0.35	Chunk overlap
bm25_k1, bm25_b	1.5, 0.75	BM25 params
rrf_dense, rrf_bm25, rrf_graph	0.75, 0.05, 0.03	RRF channel weights
retrieval_top_k	50	Candidates per channel
rerank_top_n	15	Candidates sent to reranker
final_top_n	10	Final results returned
filename_keyword_boost	2.0	Post-rerank filename tie-breaker
tier_c_penalty	0.4	Penalty for abstract base classes
context_budget_tokens	4096	Context assembly budget

API keys are read from MISTRAL_API_KEY and ZEROENTROPY_API_KEY env vars.

---

Why not just use ripgrep / the IDE's built-in search?

Lexical search finds the keyword. It doesn't find the concept. Asking "how does routing work in flask" with grep gives you every line containing "route" — most of which is irrelevant. kinetic-context returns the 8 functions that actually implement routing, with their full bodies and signatures, in 68ms.

Why not just stuff the whole repo into the LLM context window?

A 1k-file repo is ~500k tokens just for the source. That's expensive, slow, and the LLM's attention degrades badly past ~100k tokens. kinetic-context returns the 10 chunks that matter, fits in any context window, and costs 100x less to query.

Why a Code Knowledge Graph?

Because "what calls what" matters. When you ask "where is request used?", the graph says "the request proxy is defined in flask/globals.py, used in 47 places, and its setter is in flask/app.py". Pure lexical search sees the word request 500 times. The graph sees the relationship.

---

License

MIT. See LICENSE.

Contributing

Issues and PRs welcome at github.com/zai-org/kinetic-context.