GSMA/telco-retrieve-chunks

Telco-Retrieve Chunks — 3GPP Rel-19 Vector Database Grid

120 pre-built retrieval indexes over the 3GPP Release-19 corpus — one per (embedding model × chunking strategy × enrichment strategy) configuration. Each is a complete ChromaDB dense index plus a parallel SQLite FTS5 (BM25) index, ready to query with no re-ingestion.

Part of the GSMA Open Telco AI initiative. Companion to GSMA/telco-retrieve-QnA (the evaluation questions) and the open-telco-rag pipeline that reads these indexes directly.

~1.05 TB total. Individual configs range from ~1 GB to ~25 GB. Download the one config you need, not the repo.

Dataset Summary

Ingesting a large standards corpus into a vector store is the expensive, non-reproducible step in a RAG benchmark. This dataset ships that step's output for every point in a 120-cell grid so that retrieval and generation can be evaluated identically across configurations, on any machine, with no GPU and no ingestion run.

Each configuration was produced by open-telco-rag ingest 3gpp and pushed with push-hf.

The grid (6 × 5 × 4 = 120)

Axis	Values
Embedding model (6)	minilm (all-MiniLM-L6-v2, 384d) · mpnet (all-mpnet-base-v2, 768d) · e5 (intfloat/e5-large-v2, 1024d) · bge (BAAI/bge-large-en-v1.5, 1024d) · otel-109m (OTel-Embedding-109M, telecom fine-tuned, 768d) · otel-0.6b (OTel-Embedding-0.6B, telecom fine-tuned, 1024d)
Chunking strategy (5)	text_baseline (1024 char / 200 overlap) · sliding_window_tokens (512 tok / 50) · parent_child (2048 parent / 512 child) · hierarchical_markdown (2048 / 200) · lumber_chunker (LLM semantic chunking, t<theta>)
Enrichment strategy (4)	none · metadata_tagging · acronym_expansion · llm_metadata (LLM-generated per-chunk metadata)

Repository Structure

One folder per configuration, named:

3gpp-r19_<model>_<chunking>_<enrichment>_<params>/

Examples:

• 3gpp-r19_bge_text_baseline_none_c1024_o200/
• 3gpp-r19_otel-0.6b_parent_child_llm_metadata_p2048_c512/
• 3gpp-r19_minilm_lumber_chunker_none_t550/

Inside each folder:

Path	What
chroma.sqlite3	ChromaDB metadata + document store
<uuid>/data_level0.bin, header.bin, length.bin, link_lists.bin, index_metadata.pickle	HNSW dense index for the collection
<collection>_fts.db (-wal, -shm)	SQLite FTS5 BM25 index (parallel keyword search)
_ingest_complete.json	Per-config ingestion manifest — complete, files_total, files_processed, skipped_files, error_files, chunks_stored, timestamp

Dataset viewer note: Hugging Face auto-builds the viewer from the 120 _ingest_complete.json files, so it shows a 120-row ingestion-manifest table, not the chunk text. The chunk text lives inside chroma.sqlite3 / the FTS DBs and is reached through the retrieval library.

Usage

pip install "open-telco-rag[local-embed]"

# pull one configuration
open-telco-rag pull-hf GSMA/telco-retrieve-chunks \
  3gpp-r19_bge_parent_child_llm_metadata_p2048_c512 --local-dir ./db

# query it — dense, BM25, or hybrid — no ingestion needed
open-telco-rag search "handover procedure for conditional PSCell change" \
  --source-type 3gpp --db-path ./db --top-k 5

# or run the full evaluation against GSMA/telco-retrieve-QnA
open-telco-rag evaluate 3gpp --db-path ./db

Raw access without the library: open chroma.sqlite3 with the chromadb client, or query <collection>_fts.db with any SQLite client (FTS5 MATCH).

Source Data & Curation

• Corpus: 3GPP specifications, Release 19 (all series), the same underlying text as GSMA/telco-retrieve-QnA. Sourced via the TSpec-LLM dataset and the 3GPP spec archive (3gpp.org/ftp/Specs/archive).
• Pipeline: open-telco-rag — per-SDO source loader → clean → chunk (per strategy) → enrich (per strategy) → embed (per model) → ChromaStore + SQLite FTS5. See the open-telco-rag README for stage details.
• Curation rationale: a fully-crossed grid lets an evaluation isolate the effect of each axis (embedding model, chunking, enrichment) on retrieval quality, holding the corpus and question set fixed — which is the point of the companion GSMA/telco-retrieve-QnA benchmark.

Considerations & Limitations

• Scale. ~1.05 TB. Pull individual config folders; never clone the whole repo.
• Single corpus / single release. 3GPP Rel-19 only. No other SDOs, no other releases.
• Point-in-time. Reflects Rel-19 spec text as ingested in August 2026.
• Reproducibility, not novelty. These are derived indexes; the value is identical retrieval state across configs, not new source data.
• Format lock-in. ChromaDB + SQLite-FTS5 on-disk formats; readable via open-telco-rag or the chromadb / sqlite3 clients, not as plain Parquet/JSONL.
• .db-wal / .db-shm sidecar files may be present; SQLite recreates them on open, so a consumer can ignore them.

Licensing

• Index artifacts and manifests: Apache-2.0, consistent with the open-telco-rag toolkit. (Set here as apache-2.0; change if GSMA requires a different licence for this repo.)
• Stored chunk text (inside chroma.sqlite3 / the FTS DBs) consists of excerpts of 3GPP specifications. That text remains subject to 3GPP's copyright and redistribution terms and those of the upstream source (the TSpec-LLM dataset / the 3GPP spec archive). Downstream users are responsible for compliance.