Skip to content

Releases: SEBK4C/Chimera.APE

Chimera.APE v0.1.0-alpha

13 Jun 04:13
@SEBK4C SEBK4C
v0.1.0-alpha
4b152d8

Choose a tag to compare

View all tags
Chimera.APE v0.1.0-alpha Pre-release
Pre-release

Chimera.APE v0.1.0-alpha

Three queries, one binary, zero regrets. Runs on anything, answers to no one.

First alpha of Chimera.APE — one Actually Portable Executable that fuses a
llamafile + Gemma 4 12B (embeddings and chat), QLever (SPARQL graph

  • BM25 text index), and TurboVec (quantized ANN) into a single offline
    hybrid graph-vector knowledge engine. Point it at a directory of files —
    text, code, images, audio — and ask questions; answers come back with
    synthesized, cited, checksum-verified provenance. Nothing leaves your machine.

What you need

  1. chimera.ape (~315 MB) — orchestrator with all three organ binaries
    embedded (llamafile, qlever-server, qlever-index, turbovec-server).
    Download from this release.
  2. Gemma 4 12B QAT q4_0 weights + multimodal projector (sidecar,
    Apache 2.0, ~7.3 GB total) from HuggingFace: 
    HF=https://huggingface.co/google/gemma-4-12B-it-qat-q4_0-gguf/resolve/main
curl -LO $HF/gemma-4-12b-it-qat-q4_0.gguf          # ~7.1 GB
curl -LO $HF/mmproj-gemma-4-12b-it-qat-q4_0.gguf   # ~175 MB (images/audio)
    Put both in the same folder; chimera finds the mmproj- file beside the
    model automatically. (Text-only works without the projector.)
  3. Any x86_64/aarch64 Linux, macOS, or Windows box. CPU-only is supported but
    slow (~7 tok/s — minutes per document at ingest); a GPU makes it
    interactive.

Quick start

chmod +x chimera.ape

# Ingest a directory (first run unpacks the embedded organs into .chimera/runtime/):
./chimera.ape ingest ~/notes --model ./gemma-4-12b-it-qat-q4_0.gguf

# Ask:
./chimera.ape --search "what did we decide about the billing rewrite?" \
    --db ~/notes/.chimera --model ./gemma-4-12b-it-qat-q4_0.gguf
Maria Chen leads Project Phoenix [1]. It is a rewrite of the billing system [1].

Sources:
  [1] phoenix.md#1  ✓ verified

✓ verified = the cited file is byte-identical to what was ingested;
⚠ drifted / ⚠ missing tell you when it isn't. Citations are promises the
checksum keeps.

Images and audio

PNG/JPEG/WAV/MP3 are first-class documents. At ingest the model transcribes
any legible text or describes the scene/sound, indexes that derived text, and
stores the raw media embedding for query-by-example:

./chimera.ape --search "the budget figure on the banner" --db ... --model ...   # text → media
./chimera.ape --search-file query.png --db ... --model ...                      # media → everything

Other commands

./chimera.ape status  --db DIR/.chimera        # counts, dims, index staleness
./chimera.ape verify  --db ... [--paranoid]    # re-checksum the corpus
./chimera.ape vacuum  --db ...                 # purge superseded data, rebuild text index
./chimera.ape sparql  "SELECT ..." --db ...    # raw SPARQL into the graph

Acceptance (Linux x86_64, CPU)

Run via scripts/acceptance.sh against the §11 criteria of the spec:

Criterion Result
APE format, statically linked ✓
Ingest mixed corpus; re-run 100% skipped < 2 s / 1000 docs ✓
Middle-edit supersession; --paranoid catches sampled-hash blind spots ✓
Search returns ≥1 ✓ verified citation; deleted source → ⚠ missing ✓
vacuum shrinks stores; sparql walks the live graph ✓
kill -9 mid-ingest then resume: no duplicates ✓

Known alpha limitations

  • Sequential ingest. CPU extraction dominates wall-clock (~2–6 min/chunk
    at 7 tok/s). The §5 bounded-queue concurrency is designed, not yet wired.
  • Incremental text index. Ingests after the first don't extend the BM25
    text index (vector + graph search unaffected); status flags it, vacuum
    rebuilds it.
  • Platform. The orchestrator, QLever, and llamafile are true fat APEs;
    turbovec-server is Rust-in-APE with Linux ABI assumptions, so Linux
    x86_64 is the tested platform for this alpha — other OSes are
    expected-but-unverified.
  • Vision OCR of dense rendered text has a known upstream pipeline bug;
    photos/scenes describe well, screenshots may transcribe imperfectly.
  • Multimodal embeddings are CPU-only (the server refuses them on GPU);
    media search degrades to the text bridge there, which is the primary path.

Full design rationale: docs/DECISIONS.md. Usage detail: docs/USAGE.md.

Loading