nlweb-data-loading

NLWeb Data Loading

Before writing code

Fetch live docs:

1. Fetch https://github.com/nlweb-ai/NLWeb/blob/main/docs/tools-database-load.md for the canonical db_load.py reference.
2. Inspect AskAgent/python/data_loading/db_load.py and db_load_utils.py in the live repo for exact CLI flags — they've added flags in recent releases.
3. Check AskAgent/python/data_loading/rss2schema.py for how RSS items map to Schema.org Article objects.
4. Confirm the embedding provider used at ingest matches preferred_provider in config_embedding.yaml for the query side — mismatch = silent retrieval failure.
5. For partner backends, check docs/setup-snowflake.md, docs/setup-cloudflare-autorag.md, etc. for backend-specific ingest steps (some bypass db_load.py).

Conceptual Architecture

What db_load Does

db_load.py is the canonical ingest pipeline. Given a source and a site name, it:

1. Fetches the source (RSS feed, JSON-LD URL, sitemap-derived URL list, CSV).
2. Normalizes each item to a Schema.org JSON object (uses rss2schema.py for feeds; passes JSON-LD through; maps CSV columns by convention).
3. Chunks long text fields (description, body) if needed.
4. Computes embeddings via the configured embedding provider in config_embedding.yaml.
5. Writes to the write_endpoint configured in config_retrieval.yaml.
6. Tags every record with the site value so retrieval can partition.

Supported Source Types

| Source | Detection | Notes | |--------|-----------|-------| | RSS / Atom feed | URL ending .rss, .xml, /feed, or content-type | Mapped to Article Schema.org type | | Schema.org JSON-LD | URL returns application/ld+json or HTML with embedded JSON-LD | Preserved as-is | | Sitemap.xml | URL ending sitemap.xml | Crawled for child URLs | | URL list file | --url-list path.txt flag | One URL per line; each fetched and parsed for JSON-LD | | CSV | .csv extension | Column-to-Schema.org mapping by convention; see docs |

Site Partitioning

Every record carries a site field. Queries filter by site=<name> to scope retrieval. Choose site names carefully — they're user-visible in /sites and become part of the agent UX. Conventions:

• Lowercase, no spaces, hyphens or underscores
• One site per logical content domain (not per RSS feed; aggregate related feeds under one site)

Embedding Dimension Trap

The most common ingest bug: data was loaded with embedding model A (dim 1536), but at query time config_embedding.yaml points to model B (dim 768). Retrieval silently returns garbage because vector dimensions don't align — or fails entirely if the backend enforces dimension constraints. Always verify the embedding provider hasn't changed between ingest and query.

Write Endpoint Selection

db_load.py writes to one endpoint at a time — the write_endpoint in config_retrieval.yaml, or override with --database <endpoint-name>. If you need data in multiple backends, run db_load multiple times changing the write endpoint each time.

Delete and Reload

Sites can be wiped:

python -m data_loading.db_load --only-delete delete-site <site-name>

Without --only-delete, the loader does upsert by URL — re-running on the same source updates existing records but leaves stale ones. For full refresh, delete first, then load.

Batch Sizing

--batch-size N controls how many records are embedded + written per round-trip. Defaults are sane (~100). Increase for large ingests if your embedding provider rate-limit allows.

Parallel Loading

data_loading/parallel_db_load.sh runs multiple loaders concurrently across sources. Use for cold-start across dozens of feeds. Watch rate limits on the embedding provider — Azure OpenAI has aggressive throttling.

Implementation Guidance

Loading an RSS Feed

python -m data_loading.db_load https://example.com/feed.xml my-blog

Each item in the feed becomes a Schema.org Article with headline, description, url, datePublished populated from the RSS fields. Embeddings come from concatenating headline + description (verify exact field selection in rss2schema.py).

Loading Schema.org JSON-LD

For sites that already serve JSON-LD (Recipe, Product, Event, Article, Movie, etc.), point db_load at a sitemap or URL list:

python -m data_loading.db_load --url-list urls.txt my-recipes

Each URL is fetched; the embedded JSON-LD is extracted and indexed verbatim. This is the highest-fidelity ingest path — the agent gets the full schema_object back at query time.

Loading CSV

python -m data_loading.db_load products.csv my-store

CSV columns must follow the Schema.org property naming convention (or the column-mapping rules in db_load_utils.py — verify). For products, columns like name, description, url, image, offers.price, offers.priceCurrency are common.

Overriding the Write Endpoint

python -m data_loading.db_load --database azure_ai_search source.xml my-site

Useful for parallel ingest across backends, or for promoting a dev qdrant_local index to prod Azure AI Search.

Incremental Refresh Pattern

# Daily — incremental upsert (existing records updated, new added, stale left)
python -m data_loading.db_load https://example.com/feed.xml my-blog

# Weekly — full refresh
python -m data_loading.db_load --only-delete delete-site my-blog
python -m data_loading.db_load https://example.com/feed.xml my-blog

Verifying a Load

After ingest:

• curl http://localhost:8000/sites — your site should appear
• curl 'http://localhost:8000/ask?query=test&site=my-blog&streaming=false&mode=list' — should return non-empty results
• Inspect a result's schema_object field — confirm it has the Schema.org properties you expect

Backend-Specific Ingest

Some retrieval backends bypass db_load.py entirely:

• Cloudflare AutoRAG — ingest is managed by Cloudflare; you upload to R2 and AutoRAG indexes for you. See docs/setup-cloudflare-autorag.md.
• Snowflake Cortex Search — data lives in Snowflake tables; Cortex Search indexes are created via SQL. NLWeb just queries.
• Shopify MCP — no ingest; NLWeb proxies to Shopify's MCP endpoint live.
• Bing Web Search — no ingest; live web search.

Common Failures

• db_load hangs on embedding — your embedding provider is rate-limiting. Reduce --batch-size or switch provider.
• Records load but never appear in /ask — check sites: allowlist in config_nlweb.yaml; check that write_endpoint and the enabled read endpoints actually overlap.
• Loaded RSS but schema_object is sparse — RSS doesn't carry rich Schema.org metadata. Either accept it or move to JSON-LD ingest.
• Embedding dim mismatch — re-ingest with the correct provider, or change config_embedding.yaml to match what was ingested.

Always cross-check flags against the live db_load.py — argument names drift release to release.