DDC CWICR Construction Cost Database
Multilingual construction work-item and resource cost database for AI, with Qdrant semantic search and n8n estimating pipelines.
- Stars
- ★ 289
- Last updated
- 3mo ago
- License
- NOASSERTION
- Primary language
- HTML
- FA score
- 53/100 · Major gaps
At a glance
- How it runs
- Works with
- Portable with changesOpenAI APIClaude API (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A cost estimator preparing a tender types "100 m² gypsum partition with studs and skim coat" into the Telegram text bot and gets matched rates, a resource breakdown and a total price in one reply.
- Not a fit if
- Commercial users unwilling to buy the separate DDC commercial licence
- Individuals who do not want to self-host Qdrant and n8n
- Website
- datadrivenconstruction.io ↗ set on GitHub
- Maintenance
- Commits in the last 6 monthslast commit 2026-08-21latest release v0.4.0 (2026-08-21)+0.9 ★/day over 7 days
- Source review
- 53/100 · Major gaps
What does this agent do, and when should you use it?
DDC CWICR (Construction Work Items, Components & Resources) is an open cost database holding 55,719 work items and 27,672 resources across 30 regions and 23 languages, with every record laid out in an 85-field schema covering classification, rate identity, resources, labor, machinery, price variants, aggregates and mass/services. Data ships as XLSX, Parquet and CSV, plus pre-built Qdrant snapshots that already contain OpenAI text-embedding-3-large (3072-dimension) vectors for 11 language collections, so semantic retrieval works without any embedding job of your own. Four n8n workflow JSONs turn the data into working estimators: a Telegram text bot, a photo web form, a universal bot accepting text/photo/PDF, and a CAD (BIM) pipeline that reads Revit/IFC/DWG exports and produces 4D/5D estimates. Costing follows resource-based norms — labor hours, material quantities and machine time multiplied by regional prices — which keeps the technology norms portable while only prices change per market. Licensing is split: data under CC BY-NC 4.0 plus a separate DDC commercial licence, code under Apache-2.0, and you supply your own OpenAI/Gemini keys plus self-hosted Qdrant and n8n.
The repository is organized by source country — folders such as CIS-Russia-GESN-FER-TER/, Asia-China-Dinge/, Europe-Turkey-Birim-Fiyat/ and SouthAmerica-Brazil-SINAPI/ — each holding the national base as parquet and xlsx (native language plus English), a resource catalog as csv/xlsx, a PROVENANCE.md, and a markets/ directory with the base repriced via World Bank PPP and translated for 48 target markets. Qdrant collections are named ddc_{lang}_{city}; a snapshot like DE___DDC_CWICR/DE_BERLIN_workitems_costs_resources_EMBEDDINGS_3072_DDC_CWICR.snapshot loads collection ddc_de_berlin with 55,719 points. Derived tracks such as ddc_au_sydney (sourced from UK_GBP) keep rate_code, resource_code and all norms bytewise identical to the source track, replacing only prices and translatable text. The bundled n8n workflows read a Telegram message, image or PDF, have an LLM extract work items, call OpenAI text-embedding-3-large to embed them, query Qdrant for matching rates, rerank with an LLM, and write HTML/Excel/PDF estimates. The CAD pipeline runs n8n_4_CAD_(BIM)_Cost_Estimation_Pipeline_4D_5D_with_DDC_CWICR.json through RvtExporter.exe to export a Revit model to XLSX, then processes 10 stages (project detection, phase generation, element assignment, work decomposition, vector search, unit mapping, cost calculation, 7.5 CTO validation, aggregation, report generation). All workflows take credentials from a central 🔑 TOKEN node: bot_token, OPENAI_API_KEY, GEMINI_API_KEY, QDRANT_URL, QDRANT_API_KEY.
- A cost estimator preparing a tender types "100 m² gypsum partition with studs and skim coat" into the Telegram text bot and gets matched rates, a resource breakdown and a total price in one reply.
- A site manager on a renovation job photographs a room; the photo estimator web form has GPT-4 Vision identify room type and elements, estimate dimensions from reference objects, and return an HTML cost report.
- A BIM engineer exports an IFC/Revit model and runs stage 4 decomposition so element types become trades (masonry, mortar, plaster) priced through Qdrant, producing a phased 4D/5D HTML and XLS report.
- A contractor operating in several countries benchmarks the same rate across Berlin, Toronto, Paris and São Paulo by switching between ddc_de_berlin and ddc_en_toronto collections.
- An AI developer building a cost-estimation RAG assistant loads a 3072-dimension Qdrant snapshot instead of generating its own embeddings.
- A data analyst takes the ~55 MB Parquet base into an ETL/BI pipeline to study labor hours and material consumption without running any AI service.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Ports
6333- Environment variables
YOUR_TELEGRAM_BOT_TOKENOPENAI_API_KEYYOUR_OPENAI_KEYGEMINI_API_KEYYOUR_GEMINI_KEYQDRANT_URL- Config files
.envn8n_1_Telegram_Bot_Cost_Estimates_and_Rate_Finder_TEXT_DDC_CWICR.json- Install notes cover
- Docker
How do you install or deploy this agent?
The database itself needs no install — clone the repo to get the tabular data:
git clone https://github.com/datadrivenconstruction/OpenConstructionEstimate-DDC-CWICR.git
cd OpenConstructionEstimate-DDC-CWICRIf you only want the cost data, read the Parquet/XLSX/CSV files directly. For the full AI pipeline you need n8n v1.0+, a Qdrant instance, and OpenAI plus a vision model API key. On n8n 2.0 and later the Execute Command node is disabled by default and must be re-enabled, otherwise the CAD/BIM pipeline will not run:
set NODES_EXCLUDE=[] && npx n8nAlternatively create C:\Users\YOUR_USER\.n8n\.env containing NODES_EXCLUDE=[]. For Docker, set NODES_EXCLUDE=[] in the container environment. The repository ships the n8n workflow JSON and Qdrant snapshots but no Qdrant installation script — deploy Qdrant separately; a local instance normally listens on http://localhost:6333.
How do you use this agent?
First, upload a snapshot into Qdrant so the collection name matches the language collection used by the workflow:
curl -X POST "http://localhost:6333/collections/ddc_en_toronto/snapshots/upload" \
-H "Content-Type: multipart/form-data" \
-F "snapshot=@EN___DDC_CWICR/EN_TORONTO_workitems_costs_resources_EMBEDDINGS_3072_DDC_CWICR.snapshot"Second, import a workflow in n8n via Workflows → Import from File, for example n8n_1_Telegram_Bot_Cost_Estimates_and_Rate_Finder_TEXT_DDC_CWICR.json.
Third, fill in the 🔑 TOKEN node:
{
"bot_token": "YOUR_TELEGRAM_BOT_TOKEN",
"OPENAI_API_KEY": "YOUR_OPENAI_KEY",
"GEMINI_API_KEY": "YOUR_GEMINI_KEY",
"QDRANT_URL": "http://localhost:6333",
"QDRANT_API_KEY": ""
}Fourth, activate and test: send /start to the Telegram bot, or open the form URL n8n provides. The CAD/BIM pipeline additionally needs path_to_converter and project_file set, with RvtExporter.exe installed. OpenAI GPT-4o is the default model; switch by enabling one of Anthropic Chat Model2, Google Gemini Chat Model, xAI Grok Chat Model1 or DeepSeek Chat Model and disabling the rest. Reports are written to the project folder as project_YYYY-MM-DD.html and project_YYYY-MM-DD.xls.
What are this agent's strengths and limitations?
- Resource-based costing separates labor hours, material quantities and machine time from prices, so the same norm set can be repriced per region instead of rewritten.
- Ships both tabular data (XLSX/Parquet/CSV) for 55,719 work items and Qdrant snapshots with 3072-dimension vectors already computed.
- Covers 23 languages, 11 Qdrant language collections and 30 country price tracks, so cross-region benchmarking needs no translation work of your own.
- Four ready-to-import n8n workflows cover text, photo, PDF and Revit/IFC/DWG inputs, and the LLM layer is swappable across OpenAI, Claude, Gemini, Grok and DeepSeek.
- Data is CC BY-NC 4.0 plus a separate DDC commercial licence, so commercial use requires a purchase; GitHub reports the licence as NOASSERTION.
- The full pipeline depends on external services — OpenAI embeddings, a vision model and a self-hosted Qdrant — so it cannot run fully offline and incurs per-call model costs.
- On n8n 2.0+ you must manually set NODES_EXCLUDE=[] or the Execute Command node is hidden and the CAD/BIM workflow breaks.
- The README's derived-track snapshot table is truncated, and heavy resource-level parquet files plus market-edition vector snapshots are stated to arrive only in the next release.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| DDC CWICR Construction Cost Database This agent | 53 · Major gaps | Self-hosted serviceFree + model costs | ★ 289 | 3mo ago | HTML | OpenAI API |
| txtai: All-in-One AI Framework | 57 · Major gaps | Library / SDKFree + model costs | ★ 14k | 1d ago | Python | OpenAI API · Claude API |
| DecisionBox | 52 · Major gaps | Self-hosted serviceFree + model costs | ★ 123 | today | Go | OpenAI API · Claude API |
| BrainDB Memory Database | 59 · Major gaps | Self-hosted serviceFree + model costs | ★ 117 | 1mo ago | Python | Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The license and security policy describe the categories of data, code, credentials, and external services, and specify source attribution, supporting the stronger transparency and attribution scores. Workflows involve external services such as APIs, Telegram, and Qdrant, while the supplied material does not show least-privilege implementation, user confirmation, credential-handling safeguards, or rollback mechanisms. Dependency security is also only partially supported: the policy gives a reporting process but no dependency pinning or audit evidence.
The material describes data formats and dependencies such as Qdrant, n8n, and APIs, and shows a manually triggered functional-suite workflow. It does not show pinned runtime dependencies, availability guarantees, or concrete error handling. Scores for consistency, dependency availability, and failure messages are therefore limited. README claims about dataset scale, languages, regions, and workflows are not cross-checked against source code or data inventories in the supplied files.
The README describes audiences including engineers, contractors, and investors, and scenarios involving text, images, PDFs, and CAD/BIM. It also documents multiple data formats and integration paths. The data-accuracy disclaimer and third-party-service scope help establish some boundaries, but claims such as immediate usability and production readiness are not substantiated here, and no precise trigger conditions or environment compatibility matrix is provided.
The README includes a contents structure, format descriptions, quick-start guidance, examples, and a troubleshooting entry point. The license separates data, code, documentation, and reserved materials and explains commercial restrictions, justifying full marks for licensing. Naming is generally recognizable, but the supplied files do not provide a clear changelog or release history. The security policy supplies reporting channels and response targets, though the evidence for a complete named maintenance and update path is limited.
Structured cost data, precomputed vectors, workflow descriptions, and field documentation could support search, estimating, and resource breakdowns, with outputs described for several common uses. Scores are limited because the material does not demonstrate result accuracy or workflow effectiveness. Large assets, external LLM/API dependencies, and possible commercial licensing add deployment costs, with no quantified benefit or operating-cost analysis supplied.
The license states attribution and source-category requirements, the security policy describes LFS hash checks, and the README identifies fields and methodology, making some claims traceable. The supplied packet lacks provenance files, data inventories, and implementation code needed to cross-check scale and capability claims. Some coverage and methodology statements remain documentation assertions; the license does, however, clearly state that prices may be incomplete or inaccurate and are not a substitute for formal quotations.
- The license treats data and code differently: data and documentation are CC BY-NC 4.0, code is Apache-2.0, and some material is reserved; verify each file category and its terms before commercial use.
- Cost data may be incomplete or inaccurate and is not a substitute for formal bids or local professional cost publications. The supplied material does not show credential-protection, rollback, or concrete failure-handling implementations.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.4.0 v0.4.0 — 5-wave verification: 56 PASS / 0 FAIL
Release notes excerpt
v0.4.0 — 5-wave logic & quality verification, all bases A dedicated 5-wave QA pass was run against every national base (except CIS-Russia), then the found issues were fixed and the waves re-run until clean. Wave results (56 PASS / 0 FAIL / 6 documented WARNs) Wave 1 — platform import logic: every base imports 100% of its work items through the OpenConstructionERP importer with full classification (China: 8,796/11,312 with full class + documented 1,516 unnamed positions from the lost Beijing-PDF headings). …View on GitHub ↗ -
v0.3.0 v0.3.0 — all-bases native editions + mojibake + China classification
Release notes excerpt
All-bases data-quality pass Follow-up to v0.2.0 (Turkey): fixes applied across the other 7 national bases. Native editions restored to source language - BR_pt, ES_es, VN_vi regenerated so content columns read Portuguese/Spanish/Vietnamese instead of English (was up to 68% English in BR). - Result: 0% English in every native edition (BR, ES, VN, ID, IT, ZH, TR). Mojibake repaired across all bases - Repaired double-encoded (mojibake) resource names in ES/BR/IT/GR market catalogs. - Result: 0 mojibake cells across all 8 bases' market catalogs (~10M cells). …View on GitHub ↗ -
v0.2.0 v0.2.0 — Turkey (Birim Fiyat) data-quality overhaul
Release notes excerpt
Turkey (Birim Fiyat) — major data-quality release Complete overhaul of the Turkey cost base for OpenConstructionERP import. Classification hierarchy (was empty / broken) - Filled department_name / department_code / department_type (17 disciplines: İnşaat, Mekanik Tesisat, Elektrik, Karayolları, DSİ, ...). - Rebuilt section_name / section_type from real sub-chapter titles instead of item-description garbage. - Inferred section titles for all 3,240 sub-chapters (Turkish) and translated them to 25 languages. Turkish edition (was broken) …View on GitHub ↗
FAQ
Can I use this commercially?
Do I have to use OpenAI?
What services and keys does the pipeline require?
After import, the CAD/BIM nodes show question marks — why?
How do I add a new market or language?
Is DDC CWICR Construction Cost Database free?
Is DDC CWICR Construction Cost Database open source?
Is DDC CWICR Construction Cost Database safe to use?
Is DDC CWICR Construction Cost Database still maintained?
Who makes DDC CWICR Construction Cost Database?
What is DDC CWICR Construction Cost Database's official website?
Thanks — we use this to improve the page.