NERVE

Normalized Extraction and Retrieval for Vehicle/Edge-device Knowledge

NERVE is the knowledge extraction backbone of AxisX. It ingests the full Axis Communications product portfolio — 474 camera, audio, access control, and network device datasheets — and produces a structured, queryable corpus that AI agents can query with zero hallucination on product specs.

If NERVE is wrong, AxisX is wrong. If NERVE is incomplete, AxisX hallucinates. NERVE is the moat.

---

What It Does

1. Ingests raw product datasheets (PDF + scraped TXT) from the Axis corpus
2. Normalizes text — encoding repair, column-reflow, footnote stripping, section header standardization
3. Extracts structured product specs via regex-first parsing with Claude Haiku LLM fallback for complex fields
4. Enriches every record with decoded Axis naming attributes (form factor, outdoor rating, vandal resistance, LED illumination, etc.)
5. Stores results as JSON files + SQLite index for fast filtered queries
6. Serves specs to AxisX agents via a simple CLI query interface

---

Corpus Scale

Metric	Value
Source files	474 real PDFs + paired TXT files
Product types	Cameras (P/M/Q/F/C/V), Audio (D), Access Control (A/I), Body-worn (W), Network, Accessories (T/TQ/TU), Specialty
Storage	~5 MB clean text → JSON corpus + SQLite index
Ingest time	< 10 minutes for full corpus
Update cadence	~12 new products/year (monthly incremental runs)

---

Quick Start

# Install
pip install -e .

# Ingest full corpus
nerve ingest --source /path/to/raw-pdfs

# Query a product
nerve query --model "P3268-SLVE"

# Filter search
nerve search --chipset ARTPEC-8 --dlpu
nerve search --outdoor --vandal --form-factor dome
nerve search --type camera --series P32

# Corpus stats
nerve stats

With LLM enrichment (Claude Haiku)

export ANTHROPIC_API_KEY=sk-ant-...
nerve ingest --source /path/to/raw-pdfs

When ANTHROPIC_API_KEY is set, NERVE uses Claude Haiku for complex fields (ONVIF profiles, protocol lists, analytics apps, certifications) when regex confidence falls below threshold. Skip with --no-llm.

---

CLI Reference

nerve ingest --source DIR [--corpus DIR] [--limit N] [--no-llm]
nerve query  --model "P3268-SLVE" [--field chipset]
nerve search [--chipset ARTPEC-8] [--type camera] [--series P32]
             [--dlpu] [--outdoor] [--vandal] [--led] [--stainless]
             [--wireless] [--form-factor dome|ptz|panoramic|box]
             [--min-confidence 0.8]
nerve stats  [--corpus DIR]

---

Architecture

[Source Files: PDF + TXT]
        ↓
[Ingestion]     nerve/ingest.py    — format detection, TXT/PDF/ZIP readers
        ↓
[Normalization] nerve/normalize.py — ftfy encoding repair, reflow, section headers
        ↓
[Extraction]    nerve/extract.py   — regex + Claude Haiku hybrid
        ↓
[Naming]        axis_naming.py     — decode model string → form factor, extensions, flags
        ↓
[Storage]       nerve/storage.py   — JSON corpus + SQLite index
        ↓
[Query API]     nerve/query.py     — CorpusQuery class (used by CLI + AxisX)

---

Corpus Output Schema

Each product is stored as a JSON file under nerve-corpus/products/{series}/. Key fields:

{
  "nerve_id": "uuid-v4",
  "model": "AXIS P3268-SLVE",
  "model_short": "P3268-SLVE",
  "series": "P32",
  "product_type": "camera",
  "extraction_confidence": 0.94,
  "naming": {
    "product_line": "P",
    "product_line_name": "Professional",
    "form_factor": "Fixed dome",
    "resolution_name": "4K/8MP",
    "resolution_pixels": "3840×2160",
    "extensions": ["S", "L", "V", "E"],
    "extension_names": ["Stainless steel", "LED illumination", "Vandal-resistant", "Outdoor-ready"],
    "is_outdoor": true,
    "is_vandal_resistant": true,
    "has_led": true,
    "is_stainless": true,
    "is_rugged": true,
    "summary": "P-line fixed dome, 4K/8MP, Stainless steel, LED illumination, Vandal-resistant, Outdoor-ready"
  },
  "chipset": { "model": "ARTPEC-8", "dlpu": true },
  "video": { "max_resolution": "3840x2160", "codecs": ["H.265", "H.264", "AV1"], "zipstream": true },
  "power": { "poe_type": "PoE+", "max_watts": 25.5 },
  "physical": { "ip_ratings": ["IP66", "IP67"], "ik_rating": "IK10" },
  "features": { "edge_vault": true, "signed_video": true, "acap": true }
}

SQLite Index (nerve-corpus/index.db)

Fast-filter columns without loading JSON:

Column	Use
chipset, artpec_gen, dlpu	Chipset filtering
product_type, series	Product category filtering
ip_rating, max_resolution	Spec filtering
form_factor	Form factor filtering
is_outdoor, is_vandal_resistant, has_led	Naming flag filtering
is_stainless, is_wireless, is_explosion_protected, is_panoramic	Naming flag filtering
extraction_confidence	Quality filtering

---

axis_naming.py

Standalone canonical decoder for Axis model naming conventions. Shared between NERVE and AxisX — never duplicate this logic.

from axis_naming import decode_model, nl_to_naming_filters

decode_model("M4225-LVE")
# → {'form_factor': 'Fixed dome', 'is_outdoor': True, 'is_vandal_resistant': True,
#    'has_led': True, 'resolution_name': 'HDTV 1080/2MP', ...}

nl_to_naming_filters("outdoor vandal dome cameras")
# → {'is_outdoor': True, 'is_vandal_resistant': True, 'form_factor': 'dome'}

---

Stack

Component	Technology
Language	Python 3.12
PDF extraction	pdfplumber
Encoding repair	ftfy
LLM extraction	Claude Haiku (claude-haiku-4-5) via Anthropic SDK
Storage	JSON files + SQLite
CLI	click

---

Project Structure

axis_naming.py          # Shared Axis model naming decoder (NERVE + AxisX)
nerve/
  ingest.py             # Source file scanning and reading
  normalize.py          # Text cleaning pipeline
  extract.py            # Regex + LLM field extraction
  storage.py            # JSON + SQLite persistence
  query.py              # CorpusQuery class
  cli.py                # CLI entry point
nerve-corpus/
  products/             # One JSON per product, organized by series prefix
  index.db              # SQLite index for fast queries
docs/
  spec.md               # Full Software Design Document

---

Spec

Full system design is in docs/spec.md.