Joining DVF, DPE and BDNB for one address: linking a French sale to its energy diagnosis and its building

The problem: three databases, no shared key

Each of the three databases describes a different object, and that is exactly what makes the join hard.

• DVF (DGFiP)

  Describes a transfer: a date, a price, an address, one or more parcels and units. One sale often produces several rows (a flat, its cellar, its parking space), and the row id is regenerated every time the file is reloaded.

• DPE (ADEME)

  Describes a dwelling or a building at the date of a diagnosis. A block of flats carries dozens of them, one dwelling accumulates several across sales and lettings, and the method changed in July 2021.

• BDNB (CSTB)

  Describes a building, or a group of buildings, with its construction year, materials, number of dwellings and storeys. It knows nothing of sales or individual dwellings.

Addresses are not normalised the same way in the three files, the cadastral reference is filled in for only part of the DPEs and is unique only within a commune, and one sale can correspond to several diagnoses. So there is no exact join, only matches of varying reliability, which have to be documented.

How Normi does the join

1. 1. A durable identity for every sale

   The DVF row id changes on every reload, so using it as a join key means losing every match at the next reload. Normi computes a stable transaction_group_id from the date, the normalised address and the price. DPE and BDNB matches are stored on that key, and a mapping table (18,119,513 rows) bridges it to the current id exposed by the API.

2. 2. Sale → DPE: the nearest diagnosis

   For each geocoded sale, excluding bulk sales and price outliers, Normi keeps the nearest ADEME diagnosis in the same département, within 200 metres, issued between 10 years before and 1 year after the sale. One diagnosis per sale, so that a sale in a large building does not count as many times as the building has flats. Floor area and property type are not compared: that is a coverage choice, and its consequence is shown below.

3. 3. Sale → BDNB: the nearest building

   Same principle, with no date window: the nearest BDNB building within 200 metres, in the same département. The distance is stored and returned by the API as match_distance_m. Across the 17,813,826 links, the median distance is 13.3 m and 95% are closer than 49.8 m. The table is only published once département coverage and the share of links that resolve to a sale pass their integrity thresholds.

What the join makes possible

The match serves area-level statistics first, where individual errors cancel out. Four analyses follow directly, each available as a REST endpoint and as an MCP tool.

A real example: one sale in Rennes, three calls

Starting point: rue Gilles Languedoc in Rennes (35000). The responses below were captured in production on 24 September 2026, shortened but not altered. The example starts from a street rather than a house number, and DVF contains no buyer or seller names.

# 1. Every sale on the street → a transaction id (20 credits)
curl -H "X-API-Key: $NORMI_API_KEY" \
  "https://mcp.normi.fr/v1/property-history?address=rue%20Gilles%20Languedoc&code_postal=35000"

# 2. The BDNB building linked to that sale (2 credits)
curl -H "X-API-Key: $NORMI_API_KEY" \
  "https://mcp.normi.fr/v1/transactions/18109193/bdnb"

# 3. Price per m² by DPE class in the postcode (10 credits)
curl -H "X-API-Key: $NORMI_API_KEY" \
  "https://mcp.normi.fr/v1/stats/dpe-premium?code_postal=35000&type_local=Appartement"

1. The street history gives the sale's id

Twenty-two sales since 2014. The one we follow: a 4-room flat of 86.59 m², sold on 30 December 2025 at €6,045/m². The off-plan (VEFA) sales of 2022 and 2024 on the same street already point to a new development.

{
  "address_normalized": "… RUE GILLES LANGUEDOC, 35000 RENNES",
  "transactions": [
    {
      "id": 18109193,
      "date": "2025-12-30",
      "price_per_m2": 6045,
      "surface": 86.59,
      "type": "Appartement",
      "rooms": 4,
      "nature": "Vente"
    },
    {
      "id": 16417807,
      "date": "2024-11-29",
      "price_per_m2": 5908,
      "surface": 76,
      "type": "Appartement",
      "rooms": 3,
      "nature": "Vente en l'état futur d'achèvement"
    },
    {
      "id": 16409675,
      "date": "2024-11-28",
      "price_per_m2": 6453,
      "surface": 86,
      "type": "Appartement",
      "rooms": 4,
      "nature": "Vente en l'état futur d'achèvement"
    }
  ],
  "note": "Found 22 transaction(s) at this address"
}

2. The BDNB building confirms a 2023 block

{
  "building": {
    "batiment_groupe_id": "bdnb-bg-7LTJ-8KUT-9D4U",
    "annee_construction": 2023,
    "nb_logements": 21,
    "nb_etages": 6,
    "materiaux_murs": "INDETERMINE",
    "materiaux_toit": "INDETERMINE",
    "usage_principal": "Résidentiel collectif",
    "usage_niveau_1": "Résidentiel collectif",
    "dvf_nb_mutations": 5,
    "dvf_prix_m2_median": 6453,
    "match_distance_m": 8
  },
  "transaction_id": "18109193",
  "_credits": {
    "used": 2,
    "remaining": "…"
  }
}

Matched at 8 metres: a residential block of 21 dwellings over 6 levels, built in 2023. That is consistent with the off-plan sales in the history. Materials are undetermined: the BDNB does not record them yet for this recent building.

3. Price per square metre by energy class puts the sale in context

At €6,045/m², the sale sits well above every class median in 35000, as expected for a new-build flat. The table also shows why a green premium is never read raw: here, G-rated homes sell above the median per square metre. One plausible explanation, which this table alone cannot establish: in this central postcode, the least efficient homes are often small, old and well located, two factors that push the price per square metre up. Energy class is then entangled with location and size; compare like-for-like type, size and neighbourhood before drawing a conclusion.

The same question over MCP

From Claude, ChatGPT or any MCP client connected to Normi, a single request is enough. The assistant chains lookup_property_history, get_building_characteristics and analyze_dpe_price_premium, for the same 32 credits.

Find the sales on rue Gilles Languedoc in Rennes (35000). For the 4-room flat sold on 30 December 2025, give me the building's characteristics, then compare its price per m² with the median by DPE class for flats in 35000.

To connect the server: install the Normi MCP server.

Corpus figures

These figures are read from a corpus snapshot recomputed after every data load, never calculated on the fly. The DPE figure counts distinct transactions, not match pairs.

DVF rows linked to a BDNB building

17,813,826

départements covered by the BDNB match

93 / 93

median distance between the sale and the matched building

13.3 m

95th percentile of that distance

49.8 m