index-updates-and-serving
Mutate existing ColBERT indexes with IndexUpdater, persist index changes safely, coalesce updated artifacts, and expose a lightweight search API without import-time side effects.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- index-updates-and-serving
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Mutate existing ColBERT indexes with IndexUpdater, persist index changes safely, coalesce updated artifacts, and expose a lightweight search API without import-time side effects.
- disable-model-invocation
- true
02 资源文件 6
- 参考文档references/api-reference.md3.6 KB
- 参考文档references/index-update-workflows.md5.0 KB
- 参考文档references/serving-api.md3.4 KB
- 参考文档references/troubleshooting.md4.5 KB
- 脚本scripts/index_update_template.py7.3 KB
- 脚本scripts/serve_search_api.py6.9 KB
03 SKILL.md 正文
ColBERT Index Updates and Serving
Use this sub-skill when the task is to add passages to an existing ColBERT index, remove passage IDs, decide whether changes survive Searcher reloads, persist IndexUpdater changes to disk, coalesce multi-chunk index artifacts, or serve an existing index through a small JSON search API.
Route the task
- For verified constructor signatures, method semantics, reload behavior, return shapes, and package facts, read
references/api-reference.md. - For safe add/remove/persist workflows, disk mutation precautions, coalescing, backup strategy, and reload validation, read
references/index-update-workflows.md. - For Flask-style serving, explicit CLI/env configuration,
/api/searchresponse shape, lazy Searcher creation, andk <= 100behavior, readreferences/serving-api.md. - For install/import, optional dependency, backend, pid, metadata, chunk, mmap, and server configuration failures, read
references/troubleshooting.md. - For an editable add/remove/persist command template with no ColBERT import at
--helptime, adaptscripts/index_update_template.py. - For a safe lightweight JSON search service adapted from ColBERT's public server pattern with lazy Searcher creation, adapt
scripts/serve_search_api.py.
Core boundaries
- Cover
IndexUpdater(config, searcher, checkpoint=None),Searcher(...)reload behavior, index artifact mutation,persist_to_disk(), coalesced index directories, and lightweight HTTP serving. - Keep fresh index creation, baseline search,
search_all, ranking save, and general search tuning in../indexing-and-search/. - Keep TSV/qrels/ranking evaluation, query/collection validation details, and metrics workflows in
../data-and-evaluation/. - Keep checkpoint internals, tokenizer/model compatibility, and training/distillation in
../modeling-and-tokenization/and../training-and-distillation/.
Fast starting point
- Open an existing index with
Searcher(index=..., checkpoint=..., collection=..., config=..., index_root=...)and createIndexUpdater(config, searcher, checkpoint=...)when adding passages. - Use
remove([pid, ...])for in-memory removal; useadd([passage, ...])for in-memory append; callpersist_to_disk()only after backing up or writing to a disposable copy of the index. - Reload a fresh
Searcherafterpersist_to_disk()to prove changes survived disk mutation; without persistence, changes affect only the currentSearcherinstance. - Serve search through
scripts/serve_search_api.pywith explicit--index-nameand either--index-rootor a compatible ColBERT config; do not create aSearcherat module import time.
Verified package facts
- Distribution:
colbert-ai0.2.22; import package:colbert. - Verified imports include
colbert,colbert.infra,colbert.data,colbert.modeling.checkpoint,utility, andbaleen. - CPU import inspection works and CPU search can work, but practical indexing, training, and passage encoding for updates usually require CUDA/GPU-capable PyTorch.
- Important signatures:
Indexer(checkpoint, config=None, verbose=3),Searcher(index, checkpoint=None, collection=None, config=None, index_root=None, verbose=3),IndexUpdater(config, searcher, checkpoint=None),Collection(path=None, data=None),Queries(path=None, data=None), andRanking(path=None, data=None, metrics=None, provenance=None).
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).