README.md

metadata

title: SML Conversation Tree Explorer
emoji: 🌳
colorFrom: purple
colorTo: green
sdk: docker
app_port: 7860
pinned: true
short_description: Explore conversation trees with a traceable Python decoder

SML Conversation Tree Explorer

Follow recorded conversation branches in real time. Every selected response keeps its full history, original tile metadata and JSON pointer back to its source.

Open the live explorer · Public dataset and decoder · SMLBuilder

What the demo does

Choose a file and conversation, select a recorded reply, then follow its exact continuation. Rewind to any selection to explore another branch. Inspect a turn's original metadata or download a replayable trace containing the full conversation, zero-based branch indices, pinned repository revision and verified content hash. Share a link to the same file, conversation and path.

The catalog references the existing public AAC Children and TinkyBrain corpora at pinned revisions. Files load on demand; one small public starter file is bundled for immediate use. Source bytes must match their recorded Git/LFS hash before decoding. Dataset files include trees, linear chats and derived training tables; unsupported files are explicitly identified rather than turned into conversations.

The runtime has no neural inference, model weights, GPU requirement or API key. It replays recorded content; it does not validate its accuracy or turn specialist training examples into professional advice.

Lightweight decoder

tree_decoder.py is standalone Python 3.9+ with no third-party dependencies. It extends the original AAC CLI's branch walker with path-preserving history, multiple native schemas, exact source pointers, linear-chat replay and JSONL export. The original is preserved unchanged in public_decoder/AAC_cli_original.py.

python3 tree_decoder.py conversation.json --list
python3 tree_decoder.py conversation.json --conversation 0 --path 0,2,1
python3 tree_decoder.py conversation.json --all > traces.jsonl

Supported native schemas: question/branches/answer, initial_statement/response_branches/aac_response, hospital root/doctor/patient, both caller/dispatcher orientations, and recorded messages sequences. Sentence text, code examples and metadata are retained exactly. Equal text does not merge branches. Nested continuations never become separate conversation roots.

--all walks every reachable state in every recognized root and includes complete history for each path. A malformed branch raises an explicit error. An unsupported file does not silently become a successful empty export. The web interface limits paths to 128 selections and source files to 64 MiB; the standalone decoder does not apply those limits. Files can have multiple roots and independently terminating branches.

Run and verify

python3 -B -m unittest discover -s tests -v
python3 app.py

Open http://localhost:7860. The Docker Space uses Python's standard-library HTTP server on port 7860, with bounded request concurrency and a three-document LRU cache. No dependency installation is needed. Runtime state is ephemeral; visitor paths are encoded in their share links, not saved as personal conversations.

Public read-only endpoints: /health, /api/catalog, /api/document?id=…, /api/trace?id=…&conversation=0&path=0,2, and /api/export with the same parameters. Unrecognized IDs, out-of-range branches, source-integrity failures and unsupported continuations fail explicitly. Only catalog-listed public URLs can be fetched.

Provenance and publication scope

The original walker is available in the public AAC Children dataset. The catalog records exact source revisions, filenames and Git/LFS hashes. Upstream dataset terms remain applicable; this showcase does not grant additional rights to underlying material. Recovery archives, machine-path manifests, credentials and personal runtime history are excluded from this public release.