nlweb-ask-endpoint

name	nlweb-ask-endpoint
description	Implement and consume the NLWeb /ask REST endpoint — request shape (GET/POST, query-string and v0.55 structured body), SSE streaming response, modes (list/summarize/generate), in-stream "message_type" headers, error envelopes, and client-side parsing. Use when building an NLWeb server route, calling /ask from a custom agent, or debugging /ask responses.
allowed-tools	Read, Write, Edit, Bash, Grep, Glob, WebSearch, WebFetch

NLWeb /ask Endpoint

Before writing code

Fetch live spec:

1. Fetch https://github.com/nlweb-ai/NLWeb/blob/main/docs/nlweb-rest-api.md for the canonical /ask contract — request params, response shape, and streaming format.
2. Fetch https://github.com/nlweb-ai/NLWeb/blob/main/docs/life-of-a-chat-query.md to trace a request end-to-end.
3. Fetch https://github.com/nlweb-ai/NLWeb/blob/main/docs/nlweb-headers.md for the in-stream "headers" mechanism (license, data-retention, rate-limit messages are NOT HTTP headers).
4. Web-search the latest release notes for the v0.55+ structured POST body shape (query.text, prefer.mode, prefer.streaming, meta.version) — this is newer than the GET-only legacy contract.
5. Check webserver/routes/api.py in the live repo to confirm exact param names.

Conceptual Architecture

Routes

Route	Method	Purpose
/ask	GET, POST	Main NL query
/who	GET	Site relevance for a query (federated)
/sites	GET	List configured sites
/config	GET	Public config (safe subset)

Request Parameters

Verify exact names against the live routes/api.py. Stable subset:

Param	Type	Required	Default	Notes
query	string	yes	—	NL question
site	string	no	all	Backend partition; in MCP can be array
prev	string	no	—	Comma-separated previous queries (conversation context)
decontextualized_query	string	no	—	Pre-resolved query; skips server-side decontextualization
streaming	bool	no	true	"0" / "false" / "False" disables
query_id	string	no	auto	Echoed in response
mode	enum	no	list	list | summarize | generate
scorer	string	no	default	e.g., nlwebscorer for the neural reranker
itemType	string	no	—	Schema.org type hint (skip type detection)
response_format	string	no	—	v0.55 structured-body field

v0.55 Structured POST Body

The newer body format groups fields:

{
  "query": { "text": "your question" },
  "context": { "prev": ["previous q1", "previous q2"] },
  "prefer": {
    "mode": "list",
    "streaming": true,
    "response_format": "schema"
  },
  "meta": { "version": "0.55" }
}

Verify the exact field names against the live docs before relying on this — fields are still settling.

Streaming Response Format (SSE)

NLWeb uses Server-Sent Events when streaming=true (the default):

Content-Type: text/event-stream
Cache-Control: no-cache
Connection: keep-alive
X-Accel-Buffering: no

Each chunk is:

data: <json>\n\n

The <json> is one of:

• A message object (header-like): {"message_type": "license", "content": {...}}
• A partial result: {"results": [...]} (results may arrive incrementally as FastTrack streams)
• A terminal object: {"query_id": "...", "complete": true} (exact field — verify live)

In-Stream "Headers" (NLWS Mechanism)

NLWeb's "headers" are NOT HTTP response headers — they are JSON objects in the SSE stream with a message_type discriminator. Known types:

message_type	Purpose
license	Content license terms
data_retention	How long the agent may cache results
cache_policy	Caching directives
ui_component	Optional rendering hint
usage_terms	Acceptable use
rate_limits	Calls/sec / day budget
data_freshness	When the underlying data was last indexed
api_version	Server's NLWeb version

Client parsing rule: don't assume results is the first chunk. Buffer message objects until you see the result stream or a terminal marker.

Non-Streaming Response

With streaming=false, the server returns a single application/json body:

{
  "query_id": "abc-123",
  "messages": [{"message_type": "license", "content": {...}}, ...],
  "results": [
    {
      "url": "https://example.com/article/x",
      "name": "Article X",
      "site": "example",
      "score": 0.83,
      "description": "...",
      "schema_object": { "@type": "Article", "@context": "https://schema.org", ... }
    }
  ]
}

The schema_object is the original Schema.org JSON-LD that was indexed — this is what makes NLWeb results agent-actionable, not just text snippets.

Three Modes

Mode	Behavior	Use case
list	Return ranked Schema.org results, no LLM synthesis	Agent does its own rendering / re-ranking
summarize	LLM condenses top results into a short answer + still returns results	Conversational UIs
generate	Full RAG — LLM synthesizes an answer grounded in results	Q&A endpoints

Errors

For /ask, errors generally come back as 500 with a JSON envelope. For /mcp, errors use JSON-RPC 2.0:

{
  "jsonrpc": "2.0",
  "id": 1,
  "error": { "code": -32603, "message": "Internal error", "data": {...} }
}

Always check status code before parsing — partial SSE streams can drop with 200 followed by silence.

Implementation Guidance

Server-Side (extending the route)

If you need to extend /ask (e.g., add an auth check or custom param):

1. Locate webserver/routes/api.py
2. Add middleware in webserver/middleware/ rather than modifying the route directly — keeps you upgrade-safe
3. Forward to NLWebHandler (core/baseHandler.py) unchanged so the streaming + ranking pipeline still runs

Client-Side (calling /ask)

# Python streaming client (sketch — verify response shape against the live spec)
import httpx, json

async with httpx.AsyncClient() as client:
    async with client.stream("GET", "http://localhost:8000/ask",
                              params={"query": "best running shoes", "site": "shoes", "mode": "generate"}) as r:
        async for line in r.aiter_lines():
            if not line.startswith("data: "):
                continue
            obj = json.loads(line[6:])
            if "message_type" in obj:
                handle_header(obj)
            elif "results" in obj:
                handle_results(obj["results"])

When to use which mode

• Agent that re-ranks and selects on its own → mode=list
• Quick conversational answer with citations → mode=summarize
• Single synthesized answer (chatbot-style) → mode=generate

Debugging /ask

• Set streaming=false first — easier to inspect a single JSON body.
• Add decontextualized_query to bypass query rewriting and isolate ranking issues.
• Try mode=list to see the raw retrieval — if results are bad here, the problem is ingest/embeddings, not the LLM.
• Pass query_id and grep server logs for it.
• Disable tool_selection_enabled in config_nlweb.yaml to bypass the router and force straight retrieval.

Always verify the exact param names and message_type values against the live spec — they evolve.