NLWeb Retrieval Backends

Before writing code

Fetch live docs:

1. Fetch https://github.com/nlweb-ai/NLWeb/blob/main/docs/nlweb-retrieval.md for the architectural overview.
2. Fetch https://github.com/nlweb-ai/NLWeb/blob/main/config/config_retrieval.yaml for the canonical list of endpoint names and their defaults — config keys move release to release.
3. Pick the per-backend setup page from docs/setup-*.md (Qdrant, Azure AI Search, Elasticsearch, OpenSearch, Postgres, Snowflake, Cloudflare AutoRAG).
4. Inspect AskAgent/python/retrieval_providers/<backend>.py for the exact client signature and required env vars.
5. Verify the embedding dimension and metric (cosine/dot/L2) the backend expects — must match the embedding provider.

Conceptual Architecture

The Read-Fanout, Single-Write Pattern

NLWeb does something unusual: it reads from every enabled retrieval endpoint in parallel and deduplicates by URL, but writes go to exactly one write_endpoint. This means:

• You can run a hybrid index (e.g., local Qdrant for site content + Bing for fresh news) without code changes
• You migrate between backends by re-running db_load against the new write_endpoint
• "Result quality" is the union of all enabled stores — a noisy backend pollutes the top-k

All Supported Backends

Read vs Write

Every endpoint declares:

• enabled: true/false — whether /ask queries it
• read: true/false — finer-grained: enable for reads
• (only one endpoint should be the write_endpoint)

The default config has qdrant_local, nlweb_west, and shopify_mcp enabled — for local dev disable the last two.

Choosing a Backend

Embedding Dimension Compatibility

Each backend stores fixed-dimension vectors. The embedding provider must emit the same dimension:

Pick the embedding provider FIRST, configure the backend's index to match THAT dimension, then ingest.

Metric Compatibility

Most NLWeb providers use cosine similarity. When creating a new index manually (Azure AI Search, OpenSearch, Postgres) make sure the metric matches what the retrieval provider class expects. Look in retrieval_providers/<backend>.py for the metric the SDK call passes.

The nlweb_west Trap

nlweb_west is a Microsoft-hosted demo Azure AI Search instance that's enabled by default. For most users this:

• Pollutes results with MS demo data
• Requires the demo's Azure credentials to even connect
• Costs nothing but adds latency

Disable it in local dev unless you specifically want the demo content.

Implementation Guidance

Switching the Write Endpoint

Edit config/config_retrieval.yaml:

write_endpoint: azure_ai_search

endpoints:
  qdrant_local:
    enabled: false
  azure_ai_search:
    enabled: true
    api_key_env: AZURE_SEARCH_API_KEY
    endpoint_env: AZURE_SEARCH_ENDPOINT
    index_name: nlweb-main

Then re-ingest:

python -m data_loading.db_load --only-delete delete-site <site>
python -m data_loading.db_load <source> <site> --database azure_ai_search

Running Multiple Backends in Parallel

Leave several enabled: true simultaneously — /ask will fan out reads, dedup by URL. Useful for:

• Hybrid: Postgres (cheap) for site content + Bing for live web facts
• A/B: Qdrant + Azure AI Search to compare retrieval quality

Adding a New Backend Provider

If NLWeb doesn't ship the backend you need:

1. Subclass the base in retrieval_providers/ — look at any existing one for the contract (search, upsert, delete-by-site).
2. Add an endpoint entry in config_retrieval.yaml.
3. Register the class in the provider factory (verify location in current code).
4. Ingest a test site.

Backend-Specific Notes

Qdrant local: zero setup; collection lives at ../data/db. To reset, delete the directory.

Azure AI Search: create the index manually (or via the setup doc's ARM template). Vector field must be vector (or whatever the provider class names it — verify).

Postgres + pgvector: CREATE EXTENSION vector; then ensure the column is vector(1536) or whichever dim matches your embedding. NLWeb uses cosine distance by default.

Snowflake Cortex Search: data is in a Snowflake table; you create a CORTEX SEARCH SERVICE over it. NLWeb queries via the Cortex API. No db_load.py involvement.

Cloudflare AutoRAG: upload files to R2, point AutoRAG at the bucket, wire NLWeb to the AutoRAG endpoint. CF manages indexing.

Shopify MCP: zero ingest. NLWeb proxies queries to a Shopify store's MCP endpoint. Configure the shop domain per-site. Disable for non-commerce deployments.

Bing: API key required; only useful combined with at least one vector backend (Bing returns web pages, not your indexed content).

Debugging Empty Results

Diagnostic ladder:

1. curl http://localhost:8000/sites — site is registered?
2. curl 'http://localhost:8000/ask?query=test&site=X&mode=list&streaming=false' — any results at all?
3. Disable all backends except the one you wrote to — does the write_endpoint return results?
4. Check embedding dimension: python -c "from embedding_providers import get_default; print(get_default().dim)" (verify exact API) and compare to your index schema.
5. Inspect the raw store: qdrant CLI / Azure Search Studio / SELECT count(*) FROM index for Postgres.
6. Re-ingest with the embedding provider that matches the index.

Always re-fetch config_retrieval.yaml from the live repo before generating config — keys change.