A database for analyzing agent trajectories, built on Apache DataFusion. It loads ATIF traces into Parquet tables and adds functions for what plain SQL handles badly: parsing tool calls (PromQL, bash, kubectl), sequence patterns over steps, comparing runs, and LLM-labelled meaning. Inspired by https://arxiv.org/abs/2609.07782.
cargo build --release
./target/release/tdb ingest data # ATIF trajectory.json + run-*/results.csv -> db/*.parquet
./target/release/tdb sql "SELECT run, count(*), avg(n_steps) FROM sessions GROUP BY run ORDER BY run"
data/ is the HolmesGPT/SREGym trace set from ProjectASAP/ASAPWorkloads PR #10 (traces/holmesgpt-sregym), not checked in.
tdb ingest <root> |
load traces, then build materialized views |
tdb sql "<stmts>" |
run ;-separated SQL, print the last result |
tdb bench [--update] |
run queries/*.sql, diff against .golden |
tdb examples [--update] |
check every example in EXAMPLES.md against examples/fixture |
tdb synth <N> <dir> |
N× replicated corpus, for timing |
Env: TDB_DB (db dir), TDB_DEFS (startup SQL, default defs.sql), TDB_LAZY=1 (ignore materialized views), TDB_LLM_* (see EXAMPLES §6).
- EXAMPLES.md: the user guide. Tables, functions and every kind of analysis, with runnable examples.
queries/: full-corpus analyses with golden outputs (T*tool calls,W*walks and patterns,X*run comparisons,llm/LLM-based).src/:main.rs(CLI, ingest),views.sql(derived views),promql.rs/shell.rs(parsers),seq.rs(seq_match),align.rs(run alignment),llm.rs(llm()),udf.rs(CREATE FUNCTION).- PLAN.md (design and milestones), STATUS.md (what is done, findings, open issues).