docs: elaborate README with sources, schema, ingestion pipeline, namespace breakdown, and library hyperlinks

README.md

@@ -1,59 +1,234 @@
 ---
 language:
 - en
+license: other
 tags:
 - rag
 - retrieval
 - technical-docs
 - programming
+- code
+- computer-science
+- embeddings
+- chunked
 size_categories:
 - 100K<n<1M
+pretty_name: DepthAPI Technical Corpus
+dataset_info:
+  features:
+    - name: chunk_id
+      dtype: string
+    - name: source_name
+      dtype: string
+    - name: source_url
+      dtype: string
+    - name: namespace
+      dtype: string
+    - name: upstream_license
+      dtype: string
+    - name: document_id
+      dtype: string
+    - name: chunk_index
+      dtype: int64
+    - name: content_hash
+      dtype: string
+    - name: content
+      dtype: string
+    - name: token_count
+      dtype: int64
+    - name: chunker_version
+      dtype: string
+    - name: retrieved_at
+      dtype: string
+  splits:
+    - name: train
+      num_examples: 260954
 ---
 
 # depthapi_technical_corpus
 
-## Summary
+**260,954 chunks** of curated technical knowledge, purpose-built for hybrid RAG systems, embedding benchmarks, reranker evaluation, and coding agent retrieval experiments.
 
-`depthapi_technical_corpus` is a mixed-license retrieval corpus built from technical documentation, books, engineering writeups, and repository-derived reference material. It is intended for open-source RAG systems, embedding benchmarks, reranker evaluation, and enterprise retrieval experiments.
+---
+
+## Overview
+
+`depthapi_technical_corpus` is a mixed-license retrieval corpus assembled from technical books, engineering documentation, system design references, coding interview prep material, and post-mortem writeups. It is the backing knowledge store for the [DepthAPI](https://github.com/sanjeevafk/depthapi) hybrid RAG pipeline.
+
+The corpus is organized into **namespaces** that map to distinct retrieval modes, allowing downstream systems to scope queries to relevant knowledge domains.
+
+---
+
+## Data Sources
+
+| Source | Type | Namespace | License |
+|--------|------|-----------|---------|
+| Deep Learning with Python (Chollet) | Book | `default` | Manning fair-use |
+| Designing Data-Intensive Applications (Kleppmann) | Book | `default` | O'Reilly fair-use |
+| System Design resources | Docs / Markdown | `default` | Mixed / CC |
+| CPython Documentation | Official Docs | `default` | PSF-2.0 |
+| Engineering post-mortems | Blog / Web | `default` | CC-BY / Mixed |
+| [Coding Interview University – Cheat Sheets](https://github.com/jwasham/coding-interview-university) | PDF Cheat Sheets | `cs_fundamentals_knowledgeset` | CC BY-SA 4.0 |
+
+> ⚠️ This is a **mixed-license dataset**. Downstream users must inspect the `upstream_license` field and `SOURCES_MANIFEST.yaml` before redistribution or commercial use.
+
+---
+
+## Namespaces
+
+Namespaces act as logical partitions for routing retrieval queries. Each namespace is designed for a different knowledge vertical:
+
+| Namespace | Purpose | Approx. Chunks |
+|-----------|---------|----------------|
+| `default` | General ML, systems, engineering, and coding knowledge | ~259,800 |
+| `cs_fundamentals_knowledgeset` | CS fundamentals — DS&A, Big-O, OOP, STL, OS, design patterns | ~1,145 |
+
+---
 
-## Ingestion Process
+## Ingestion Pipeline
 
-The corpus was assembled from multiple upstream source types and normalized through a reproducible local pipeline backed by Supabase.
+The corpus was assembled from multiple upstream source types and normalized through a reproducible local pipeline backed by [Supabase](https://supabase.com) with pgvector.
 
-- `Scrapling` was used for live technical documentation crawling and structured HTML extraction from documentation websites.
-- [`D4Vinci/Scrapling`](https://github.com/D4Vinci/Scrapling) was used for live technical documentation crawling and structured HTML extraction from documentation websites.
-- [`opendataloader-project/opendataloader-pdf`](https://github.com/opendataloader-project/opendataloader-pdf) was used for PDF extraction when ingesting book-like and document-style technical sources into normalized markdown/text blocks.
-- Source material was then normalized, rechunked deterministically, deduplicated, validated, and exported to parquet for Hugging Face datasets compatibility.
+### Tools Used
+
+- **[Scrapling](https://github.com/D4Vinci/Scrapling)** — Stealth web crawler used for live documentation scraping and structured HTML extraction. Handles anti-bot mechanisms and extracts clean text from technical documentation sites.
+- **[opendataloader-pdf](https://github.com/opendataloader-project/opendataloader-pdf)** — PDF-to-Markdown conversion library used for extracting content from book-like and document-style PDF sources. Outputs clean Markdown blocks with layout awareness.
+
+### Processing Steps
+
+1. **Extraction** — Raw content pulled from upstream sources via Scrapling (web) or `opendataloader_pdf.convert()` (PDF).
+2. **Noise filtering** — OCR artifacts, page numbers, and garbled text filtered via regex heuristics before chunking.
+3. **Chunking** — Text split semantically via `BaseIngestor.split_text_semantic(chunk_size=800)` with configurable overlap.
+4. **Deduplication** — SHA-256 content hash (`content_hash`) deduplication. Duplicate chunks are skipped on upsert.
+5. **Validation** — Minimum token count and quality threshold enforced before persistence.
+6. **Persistence** — Chunks stored in local Supabase `knowledge_chunks` table with full metadata columns.
+7. **Embedding backfill** — Dense embeddings computed via `BAAI/bge-base-en-v1.5` (768-dim) and stored in the `embedding` pgvector column.
+8. **Export** — Parquet shards generated via `scripts/release/export_to_hf.py` and published to this repository.
+
+---
 
 ## Schema
 
-Each chunk includes:
+Each row in the dataset represents one chunk:
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `chunk_id` | `string` | UUID primary key |
+| `source_name` | `string` | Human-readable source title (e.g. `"Deep Learning with Python"`) |
+| `source_url` | `string` | Canonical upstream URL for the source document |
+| `namespace` | `string` | Retrieval namespace / knowledge domain |
+| `upstream_license` | `string` | SPDX license identifier or description |
+| `document_id` | `string` | Groups all chunks belonging to the same source document |
+| `chunk_index` | `int` | Ordinal position of chunk within its document |
+| `content_hash` | `string` | SHA-256 of chunk content — used for deduplication |
+| `content` | `string` | The raw chunk text |
+| `token_count` | `int` | Approximate token count (cl100k_base tokenizer) |
+| `chunker_version` | `string` | Version tag of the chunker that produced this chunk |
+| `retrieved_at` | `string` | ISO-8601 timestamp of when the chunk was ingested |
 
-- `chunk_id`
-- `source`
-- `source_url`
-- `upstream_license`
-- `document_id`
-- `chunk_index`
-- `retrieved_at`
-- `chunker_version`
-- `content_hash`
-- `content`
+---
 
 ## Usage
 
+### Load the full dataset
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("sanjeevafk/depthapi_technical_corpus", split="train")
+print(ds[0])
+```
+
+### Stream for large-scale use
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset(
+    "sanjeevafk/depthapi_technical_corpus",
+    split="train",
+    streaming=True,
+)
+for row in ds.take(5):
+    print(row["source_name"], "|", row["namespace"], "|", row["content"][:120])
+```
+
+### Filter by namespace
+
 ```python
 from datasets import load_dataset
 
-dataset = load_dataset("sanjeevafk/depthapi_technical_corpus", split="train", streaming=True)
-for row in dataset.take(3):
-    print(row["chunk_id"], row["source"], row["upstream_license"])
+ds = load_dataset("sanjeevafk/depthapi_technical_corpus", split="train")
+
+# Get only CS fundamentals chunks (Coding Interview University cheat sheets)
+cs_chunks = ds.filter(lambda x: x["namespace"] == "cs_fundamentals_knowledgeset")
+print(f"CS fundamentals chunks: {len(cs_chunks)}")
+```
+
+### Filter by source
+
+```python
+ciu_chunks = ds.filter(
+    lambda x: "Coding Interview University" in (x["source_name"] or "")
+)
 ```
 
-## Retrieval Benchmark Example
+### Build a simple BM25 retriever
+
+```python
+from datasets import load_dataset
+from rank_bm25 import BM25Okapi
+
+ds = load_dataset("sanjeevafk/depthapi_technical_corpus", split="train")
+corpus = [row["content"] for row in ds]
+tokenized = [doc.split() for doc in corpus]
+bm25 = BM25Okapi(tokenized)
+
+results = bm25.get_top_n("binary search tree time complexity", corpus, n=5)
+for r in results:
+    print(r[:200])
+```
 
-Use `data/research_corpus/benchmarks/queries.jsonl` with your retriever and score against `qrels.jsonl`.
+---
 
 ## Licensing
 
-This is a mixed-license dataset. Downstream users must inspect `upstream_license` and `SOURCES_MANIFEST.yaml` before redistribution or commercial use.
+This is a **mixed-license dataset**. Each chunk carries its own `upstream_license` value. Downstream users are responsible for verifying license compatibility before use.
+
+| License | Sources |
+|---------|---------|
+| `CC BY-SA 4.0` | Coding Interview University (CIU) cheat sheets |
+| `PSF-2.0` | CPython documentation |
+| Manning fair-use | Deep Learning with Python, DDIA |
+| Mixed / Unknown | Engineering blogs, post-mortems |
+
+See `SOURCES_MANIFEST.yaml` (included in this repository) for full per-source provenance.
+
+---
+
+## Export Workflow
+
+This dataset is published from the [DepthAPI](https://github.com/sanjeevafk/depthapi) pipeline using `scripts/release/export_to_hf.py`.
+
+**Incremental update** (recommended after new ingestion runs):
+```bash
+python scripts/release/export_to_hf.py \
+  --hf-repo-id "sanjeevafk/depthapi_technical_corpus" \
+  --append \
+  --commit-message "Add <source> chunks"
+```
+
+**Full rebuild** (replaces all shards — use after major corpus restructuring):
+```bash
+python scripts/release/export_to_hf.py \
+  --hf-repo-id "sanjeevafk/depthapi_technical_corpus" \
+  --force-rebuild
+```
+
+A watermark is saved after each export to `data/hf_export/last_export_watermark.json`. The `--append` mode reads this to push only net-new chunks.
+
+---
+
+## Citation
+
+If you use this dataset, please cite the upstream sources where applicable. The corpus is provided as-is for research and development use.