nps_prose

Domain

Outdoor Recreation +1

recreation

Backend

Elasticsearch

nps_prose-kb

LLM

gpt-4o-mini

text-embedding-3-small

Features

Incremental

KB Backend

Embedding Model

LLM Model

Chat Name

Dataset ID

Chat ID

Similarity Threshold

Keywords Weight

Primary Domain

Secondary Domains

Hold Ctrl/Cmd to select multiple

Spatial

City/state resolution

Input Guardrails (trigger words → action)

Output Guardrails (checks → actions)

KB Website URLs (one per line)

Structured Data Sources

No structured data sources configured

Unstructured Sources 0

No unstructured data sources configured

Scrape Depth

Enrichment Model

Batch Size

Facility Data

Booking URL

Booking Provider

Chunk Tokens

Top-K / Top-N

Structured Data Sources 0

{
  "_comment": "Template config. Edit: client, domain, display_name, es_kb_index, output paths. Keep enrichment+eval+indexing defaults \u2014 they're the cost/perf-tuned settings from ridb2026.",
  "_created": "2026-05-24",
  "_purpose": "Prose-rich corpus (Wikipedia national-park articles) for stress-testing ES + LLM lane independently of the structured cascade. Track=unstructured: no structured_ingest, no input_mapping mining, no parent_scope generation. Pipeline runs enrich + index_es and a gold benchmark of hand-written prose questions to measure pure-RAG quality on RIDB-adjacent narrative content.",
  "_track_note": "This client is declared as track=unstructured. pipeline/track_defaults.py fills any unspecified knobs from the 'unstructured' preset (enrichment, grader, benchmark_generation, etc.). Explicit keys in this file always override the preset.",
  "benchmark": {
    "lane": "gold"
  },
  "benchmark_generation": {
    "strategies": [
      "same_type_compare",
      "keyword_aggregation",
      "interaction_questions",
      "multi_entity_combo"
    ],
    "validation": {
      "enabled": true,
      "fail_loud": false,
      "max_strategy_fail_rate": 0.3,
      "min_pass_rate": 0.7,
      "min_scored_tests": 50
    }
  },
  "chat_proxy": {
    "enabled": false,
    "strip_citations": true,
    "tenant_id": "",
    "url": ""
  },
  "client": "nps_prose",
  "config_provenance": {
    "fail_loud": false
  },
  "coverage_audit": {
    "enforced": false
  },
  "crawl_depth": 2,
  "display_name": "NPS Prose (Wikipedia)",
  "domain": "outdoor_recreation",
  "domain_enrich": {
    "enabled": true
  },
  "embedding": {
    "model": "text-embedding-3-small"
  },
  "enrichment": {
    "_comment": "batch_size=20 is the cost/perf sweet spot for haiku 4.5 on ~20K docs. Don't switch taxonomy_model to sonnet \u2014 it's 5x more expensive and the per-call quality gain isn't worth the latency.",
    "batch_size": 20,
    "cluster_chunks": true,
    "comparison_chunks": true,
    "eligibility_chunks": true,
    "mode": "llm",
    "model": "claude-haiku-4-5-20251001",
    "policy_chain_chunks": true,
    "restriction_chunks": true,
    "taxonomy_model": "claude-haiku-4-5-20251001",
    "workers": 20
  },
  "es": {
    "force_hybrid": true,
    "kb_index": "nps_prose-kb"
  },
  "es_kb_index": "nps_prose-kb",
  "eval": {
    "analysis_model": "claude-haiku-4-5-20251001",
    "confidence_threshold": 0.6,
    "delay": 1,
    "eval_model": "claude-haiku-4-5-20251001",
    "index_wait": 10,
    "max_iterations": 5
  },
  "extraction": {
    "discover_schema": true,
    "max_tokens_per_chunk": 1200,
    "model": "claude-haiku-4-5-20251001",
    "provider": "anthropic"
  },
  "facility_data_source": null,
  "fact_pass_threshold": 0,
  "genclair": {
    "enabled": true
  },
  "grader": {
    "deterministic_first": false,
    "trust_tier_sources": []
  },
  "incremental": {
    "auto_detect": true,
    "enabled": true,
    "on_change": "reprocess",
    "on_delete": "keep"
  },
  "indexing": {
    "_comment": "Unstructured-track tuning: prose_embed_max_chars=4000 lifts source-prose embedding window. Per-doc opt-in via metadata.source_role; entity cards unaffected.",
    "bulk_size": 500,
    "checkpoint_every": 5000,
    "embedding_batch_size": 100,
    "embedding_concurrency": 4,
    "max_docs_per_run": 0,
    "prose_embed_max_chars": 4000,
    "role_filter": "qa_allowed"
  },
  "input_dir": "output_nps_prose/input/unstructured",
  "input_mapping": " ",
  "kb_backend": "elasticsearch",
  "kb_index": "nps_prose-kb",
  "kb_urls": [],
  "llm": {
    "max_tokens": 2048,
    "model": "gpt-4o-mini",
    "provider": "openai",
    "temperature": 0,
    "top_p": 0.3
  },
  "output_dir": "output_nps_prose",
  "pattern_dedup_max": 25,
  "relation_index": {
    "_comment": "Disable for pure-structured clients with > 10K entities. Triplet relationships are already deterministic via structured_input_adapter, and the LLM-based relation_index step burns 30+ min for ~zero useful inter-entity signal.",
    "enabled": false
  },
  "retrieval": {
    "_comment": "Unstructured-track tuning (2026-05-25): semantic-leaning hybrid with prose-aware boosts. knn_boost 3.0 raises kNN weight vs BM25. Entity-name boosts reduced (50/10 \u2192 20/5) so prose chunks aren't crowded out. NEW signals \u2014 tags 3.0, important_kwd 2.0 \u2014 let enrichment-generated tags finally reach the retrieval query. source_prose 4.0 favors paragraph chunks for narrative questions. validation_verified 1.0 is a soft trust-tier nudge.",
    "keyword_search": true,
    "keywords_similarity_weight": 0.3,
    "max_tokens": 2000,
    "query_expansion": {
      "enabled": true,
      "expand_ragflow": false,
      "expand_structured": true,
      "max_loops": 2
    },
    "similarity_threshold": 0.1,
    "top_k": 20,
    "top_n": 12,
    "knn_boost": 3.0,
    "gate_content_rank": true,
    "boosts": {
      "entity_name_keyword": 20.0,
      "entity_name": 5.0,
      "content": 2.0,
      "question_tks": 1.0,
      "tags": 3.0,
      "important_kwd": 2.0,
      "source_prose": 4.0,
      "validation_verified": 1.0
    }
  },
  "scheduled_ops": {
    "alert_on_regression": true,
    "alert_threshold": -2,
    "chat_id": "",
    "dataset_id": "",
    "enabled": false,
    "eval_on_fail": true,
    "scrape_interval_hours": 168,
    "test_file": "",
    "test_interval_hours": 24
  },
  "secondary_domains": [
    "recreation"
  ],
  "spatial": {
    "data_source": "pipeline/spatial_data/national/us_cities.json",
    "enabled": false,
    "resolve_cities": true
  },
  "structured_data": [],
  "structured_docs_dir": " ",
  "structured_input_dir": " ",
  "test_generation": {
    "llm_answers": true,
    "max_questions": 150,
    "min_questions": 60,
    "pattern_dedup": true,
    "pattern_dedup_max": 4
  },
  "track": "unstructured",
  "validation": {
    "enabled": true
  }
}

🚪 Pipeline Stage Gates

🎮 Pipeline Controls

⚙️ Advanced Options

Force re-run completed stages

Start from stage:

Systematic Foundation Discovery Step 0

RootFinder analyzes your document corpus to discover foundational types, tier assignments, and protected documents before extraction begins. This ensures proper ontology alignment.

Status

Not Started

No foundation analysis performed

Outputs

No outputs generated

GenClair Zone Discovery Step 1

GenClair analyzes your document corpus to discover knowledge zones using embedding clustering and domain-fit validation. This creates zone boundaries for focused extraction.

Status

Not Started

No zone discovery performed

Zones Discovered

No documents selected

Domain Profile outdoor_recreation

Entity Types (15)

Park A state park, national park, recreation area, or forest

Campground A campground, camping area, or overnight facility

Site A specific campsite, cabin, or lodging unit

Trail A hiking trail, bike path, or water trail

Activity An outdoor activity (fishing, hunting, boating, skiing, tour

Pass A day pass, vehicle permit, annual pass, or reservation

Policy A park rule, camping regulation, or safety requirement

Fee A camping fee, pass price, rental cost, or reservation fee

Amenity A campground amenity (showers, hookups, firewood, boat launc

Season A seasonal period when facilities are open or activities ava

Program A class, lesson, course, or structured activity

Facility A building, room, pool, court, or gym space

Contact A phone number, email, office, or staff person

Person A person type or group (member, guest, minor, senior, alumni

Schedule Operating hours, class times, or semester dates

Relationships (16)

COSTS Entity has a fee or price

LOCATED_IN Entity is at a park or area

CONTAINS Park contains campgrounds, trails, or facilities

OFFERS_ACTIVITY Park or area offers an activity

HAS_AMENITY Campground or site has an amenity (electric, water

GOVERNED_BY Activity or site is subject to a rule or regulatio

REQUIRES Activity requires a permit, license, reservation,

RESTRICTS Policy restricts an action (no fires, no pets, qui

AVAILABLE_IN Activity or facility is available during a season

ACCESSIBLE_VIA Site is accessible via a road, trail, or water rou

NEAR Entity is near another entity (lake, town, trailhe

HAS_CAPACITY Site or campground has a capacity or number of uni

OFFERS Facility or organization offers a program or servi

ELIGIBLE_FOR Person type is eligible for a program or facility

CONTACTS_VIA Entity can be reached through a contact method

HAS_SCHEDULE Entity operates on a schedule or time period

Enrichment

Generate Comparison Chunks

Generate Eligibility Matrix

Generate Restriction Summary

Generate Policy Chains

Safety Critical

14 question templates

Guardrails

Critical redirect rules

1 input checks

3 output checks

2 data checks

Secondary domains: recreation

Step 0: Scrape ? Pending

Site Scrape

No scraped pages yet.

Input Dir: output_nps_prose/input/unstructured

Class Scrape

No class data yet.

Step 1: Extract ? Pending

No extraction data yet. Run pipeline Step 1 (extract).

Step 2: Build ? Pending

No build data yet. Run pipeline Step 2 (build).

Step 3: Enrich ? Pending

No enrichment data yet. Run pipeline Step 3 (enrich).

Step 4: Policies ? Pending

No policy data yet. Run pipeline Step 4 (policies).

Step 5: Lookups ? Pending

No lookup data yet. Run pipeline Step 5 (lookups).

Steps 6-8: Upload / Parse / Configure ? Pending

Not uploaded yet. Run pipeline Steps 6-8 (upload/parse/configure).

Document Registry

Click to load document status

-

Total Tracked

-

New

-

Changed

Upload Documents

File	Status	Processed	Steps

Entities & Relationships

Entity Roles

-

Core

-

Supporting

-

Noise

-

Uncertain

Domain - Core Types -

Entity	Type	Role	Score	Cluster	Compare	QA	Heal	Reason

Suppression Log

Entities blocked from one or more synthesis steps. All remain searchable via retrieval.

Entity	Type	Role	Blocked From	Reason

Test Questions

Test File

Test History

Latest Results

Run Eval & Patch

Full Loop

Failures Only

Test File

Max Iterations

ES backend — no Chat ID or Dataset ID needed. Eval uses config settings automatically.

Admin Review Queue

No pending review items.

Scheduled Operations

Enable Scheduled Ops

Scrape Interval (hours)

Test Interval (hours)

Auto-eval on failures

Alert on regression

Alert Threshold (score drop)

Test File

ES backend — IDs not needed

Cron Setup

Add this line to your crontab (crontab -e):


        0 */12 * * * cd /path/to/project && python pipeline/scheduled_ops.py --config clients/nps_prose/config.json

Pipeline Log

No pipeline log found.

Query Traces 0

No query traces found. Traces are captured when questions are routed through the RetrievalRouter.

Repair Ledger 0

No repair ledger entries. Run Eval & Patch to populate.

System Health

Repair Stats

Total repairs0

Positive0

Negative0

Neutral0

Query Performance

Traces captured0

Avg latency0ms

Data Freshness

No structured data sources loaded.

No risk report available. Run the pipeline to generate failure predictions.

Pipeline UI

Backend & Retrieval

Domain & Scope

Guardrails Domain defaults

Data Sources

Unstructured Data Sources Documents, files, content

Pipeline Settings

nps_prose — config.json

🚪 Pipeline Stage Gates

🎮 Pipeline Controls

⚙️ Advanced Options

🚀 Pipeline Execution

Status

Outputs

Process Output

Foundation Review

Status

Zones Discovered

Discovery Progress

Discovered Zones

Selected Documents

Entity Types (15)

Relationships (16)

Enrichment

Guardrails

Site Scrape

Class Scrape

Document Registry

-

-

-

Entities & Relationships

Entity Types

Filter

Entities

Relationships

Entity Roles

-

-

-

-

Suppression Log

Test Questions

Generated Questions

Test History

Test Output

Latest Results

Pipeline Control Center

Data Sources Configuration

Structured Data

Unstructured Data

RootFinder Setup

GenClair Discovery

Extract Data

Foundation & Discovery

Build & Validate

Domain & Policies

Index & Test

Healing

Run Eval & Patch

Admin Review Queue

Scheduled Operations

Cron Setup

Pipeline Log

Query Traces 0

Repair Ledger 0

System Health