Table of Contents
To search decades of well reports with a private LLM, do not just load PDFs into a chatbot. Turn each report into structured records (well, date, depth, activity code, duration) for exact queries and text passages for meaning-based search. An on-premise model then combines both, cites the report and page, and keeps the data on your servers.
Key takeaways
- A daily drilling report is semi-structured: in the public Volve dataset, each file covers 24 hours with timestamped, coded activities, fluid measurements and survey stations, plus free-text comments (Lu, 2026).
- The free text is hard to search. Researchers working on 9,670 daily reports from 303 wells cite technical symbols, mistyped or abbreviated terms and incomplete sentences (Hoffimann et al., 2017).
- Codes and names change across decades. IADC's DDR Plus moved from 28 legacy codes to 34 operational codes in 2019 (Drilling Contractor), and one wellbore can appear under three different names.
- Numbers such as NPT hours should come from database queries over activity records; the language model chooses the query and explains the result.
- In Algeria, upstream data are the exclusive property of the State and managed by ALNAFT (Law 19-13, Article 18). Check with counsel before indexing them or sharing them with a vendor.
What is in a well report archive?
An operator or drilling contractor that has worked a field for thirty years holds several families of reports, produced by different teams and stored in different places. The table lists the usual ones and how each should be indexed.
| Report family | What it holds | Usual form in older archives | How to index it |
|---|---|---|---|
| Daily drilling report | 24-hour summary: activities with times and depths, codes, mud and survey data, remarks | Paper, scans, spreadsheets; XML in recent years | One structured row per activity, plus the remarks as text |
| Daily completion and workover reports | Same pattern for completion and intervention work | Paper, scans, spreadsheets | Same as drilling, with the operation type as metadata |
| End-of-well report | Narrative and tables summarising the well: problems, lessons, final status | Long PDF or scan | Split by section; keep tables whole |
| Well test reports | Rates, pressures, test conditions and interpretation | PDF with tables | Values to a table; interpretation as text |
| Mud logs and geological reports | Lithology, shows, formation tops | Scans, log images, PDFs | Formation tops as data; descriptions as text |
The daily report is the backbone. Energistics, which maintains the WITSML standard, describes its DrillReport object as designed "to capture required and some supplemental information for daily drilling reports", which operators must share with partners and government agencies (Energistics). Recent reports may therefore already be structured XML. Older ones are usually paper or scans.
Why are well reports harder to search than manuals?
A maintenance manual is written once, edited and approved. A daily report is written at the rig, every day, often at the end of a long shift. Four problems follow.
The free text is rough. The authors of one study classified sentences from 9,670 daily reports written by drilling engineers in a field of 303 oil wells. They list the main obstacles as frequent technical symbols, mistyped or abbreviated technical terms, and incomplete sentences (Hoffimann et al., 2017). A search for "pulled out of hole" will miss "POOH", and an embedding model may not know what either means.
Codes changed over time. The IADC daily drilling report, or tour sheet, carried 23 operational drilling codes and five completion codes. DDR Plus, launched in 2019, kept them, added six more, including nonproductive time and well control, and added activity, sub-activity, equipment and sub-equipment levels below each main code (Drilling Contractor). Company code lists differ again. A question across three decades needs one mapped code list.
Well names are not stable. In a 2026 study on the public Volve data, the same wellbore appeared as "15_9_F_11_T2" in report file names, "NO 15/9-F-11 T2" in WITSML headers and "15/9-F-11" in production data (Lu, 2026). Without a master list of aliases, a search on one well silently misses part of its history.
Scans and languages. Older reports exist only as scans, and OCR errors land on depths and values. If your archive mixes French, English and Arabic, Arabic is the weakest point: on the KITAB-Bench benchmark the best model reached only 65% accuracy converting Arabic PDFs to Markdown, with numeral recognition and table structure among the recurring errors (Heakl et al., 2025). Our guides to OCR for scanned technical documents and documents that mix Arabic, French and English cover both in detail.
Why not paste all the reports into the model?
Because they do not fit, and models read long inputs unevenly. Anthropic notes that a knowledge base smaller than 200,000 tokens, about 500 pages, can be placed whole in the prompt (Anthropic). One small North Sea field, Volve, produced 1,759 daily drilling report files (Lu, 2026), and Equinor's full Volve release holds about 40,000 files from a field that produced from 2008 to 2016 (Equinor). Thirty years of a large field is far beyond that. Long inputs also lose information: models use what is at the start or end of the context more reliably than what is in the middle (Liu et al., 2023).
The answer is retrieval-augmented generation (RAG): search the archive first, then give the model only the passages that matter (Lewis et al., 2020). Our RAG guide for maintenance manuals describes the general pipeline. Well reports need one addition: a structured store next to the text index.
What architecture works for well reports?
Most questions about wells mix numbers and narrative. "How many hours of stuck pipe did we have in 12¼-inch sections since 2000?" is a count. "What did crews do when they lost circulation in that interval?" is a reading task. A setup that handles both keeps two stores and lets the model call each one as a tool.
The 2026 TADI study on Volve follows this pattern: 12 structured tables with 65,447 rows in DuckDB for queries, 36,709 embedded documents in ChromaDB for semantic search, and a language model that calls 12 specialised tools to cross-reference measurements with report text (Lu, 2026). Its author concludes that tool design, more than model size, drove answer quality. Note that the study used a hosted commercial model and embeddings through an API; a private deployment replaces both with models running on your own servers.
| Question type | Example | Best route |
|---|---|---|
| Count or duration | NPT hours by cause, per well, per year | SQL over the activity table; the model writes and explains the query |
| Exact identifier | All reports mentioning a fish or a code | Keyword search on the text, filtered by well and date |
| Narrative | How a past kick or loss was handled | Semantic search over remarks and end-of-well sections |
| Combined | Wells where losses followed a mud weight change | Query for candidates, then read the matching remarks |
Keyword search matters for codes and part numbers: Anthropic gives the example of an error code that an embedding model misses but BM25 keyword search finds exactly (Anthropic). For the SQL side, apply the safeguards in our post on text-to-SQL on production databases: read-only access, curated views and the query shown with every answer.
The language model can also help build the structured store. KCA Deutag trained a classifier on 40,000 manually tagged rows of free-text drilling comments, from a dataset of 3.1 million rows, to sort each one into rig problem, well problem, normal operation, HSE event or waiting time, with more than 90% accuracy reported (Drilling Contractor). An LLM can propose the same kind of labels for old remarks that were never coded, as long as engineers check a sample before the labels feed any statistic.
How do you build it, step by step?
- Inventory the archive. Count reports by field, decade, format (XML, spreadsheet, native PDF, scan) and language, and name an owner for each source.
- Build a well master list. One identifier per wellbore, with every alias, sidetrack and renaming. A knowledge graph of plant equipment applies the same idea to tags.
- Load the digital reports first. Parse XML and spreadsheets into activity, fluid and survey tables, keeping the source file and line for each row.
- OCR the scans, then check depths, dates and values on a hand-typed sample before trusting them.
- Map activity codes from each era and company list to one reference list, and keep the original code next to the mapped one.
- Chunk the text by activity or by report section, tagged with well, date, depth, phase, code, language and page.
- Write a glossary of abbreviations and local terms, and use it to expand questions and tag passages.
- Index for both keyword and semantic search, and filter results by each user's permissions at retrieval time, as OWASP recommends with "permission-aware vector and embedding stores" (OWASP LLM08:2025).
- Test with real questions from drilling and well engineers, including questions the archive cannot answer, and measure as described in measuring hallucination in an industrial assistant.
Who owns the data, and where may it be processed?
In Algeria, Law 19-13 of 11 December 2019 defines "data" as all results, information, cores and studies, and geological, geophysical, hydrogeological and geochemical information acquired during upstream activities, before or after the law came into force, together with related technical data (Law 19-13, Article 2). Article 18 states that these data are the exclusive property of the State, that ALNAFT manages, preserves, uses, makes available and distributes them, and that the national company, contracting parties and prospectors must transmit them to ALNAFT. Article 19 lets ALNAFT set the conditions of access to its data bank.
Well reports contain much of what this definition covers. That does not by itself forbid indexing your own archive, but it means the reports are not simply the operator's to send to an outside service. This is a reading of the text, not legal advice: confirm with counsel, and with ALNAFT where relevant, before a project starts. Our oil and gas AI data readiness checklist covers the other ownership checks. Running every component on your own infrastructure, including OCR, embeddings and logs, keeps the question simple; for isolated sites, see air-gapped AI deployment.
What we have seen at KABAS AI
DALILE is our on-premise AI knowledge assistant. It reads technical documents, spreadsheets and databases, answers in plain language with exact sources, queries databases from natural-language questions, and runs on the client's own servers. That combination of document search and database queries is what a well report archive needs.
We have not published a well report search project. Our published oil and gas work is at ADFC (Sonatrach Group), where Virtual Lab supports drilling fluid formulation. The case study describes formulation knowledge that was "dispersed in various formats and locations", and ADFC is exploring integration with DALILE for knowledge discovery, literature review and technical documentation.
Our practical advice for a first project:
- Start with one field and one decade of digital reports before touching the scans.
- Pick three questions engineers ask today, such as offset well problems before a new well, and build the test set from them.
- Keep numbers in the database. Let the model explain, cite and query, not add up.
Our agentic AI service deploys DALILE and connects it to document stores and databases. See our oil and gas solutions for other use cases.
Sources
- Hoffimann, Mao, Wesley and Taylor, Sequence Mining and Pattern Analysis in Drilling Reports with Deep Natural Language Processing (arXiv 1712.01476, 2017)
- Lu, TADI: Tool-Augmented Drilling Intelligence via Agentic LLM Orchestration over Heterogeneous Wellsite Data (arXiv 2605.00060, 2026)
- Equinor, Volve field data sharing
- Energistics, WITSML Drill Report data object
- Drilling Contractor, IADC launches DDR Plus to enable more granular data collection (March 2019)
- Drilling Contractor, KCA Deutag deploys artificial intelligence to extract lost-time insights from daily drilling reports (June 2021)
- Heakl et al., KITAB-Bench: A Comprehensive Multi-Domain Benchmark for Arabic OCR and Document Understanding (arXiv 2502.14949, 2025)
- Anthropic, Introducing Contextual Retrieval (September 2024)
- Liu et al., Lost in the Middle: How Language Models Use Long Contexts (arXiv 2307.03172, 2023)
- Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (arXiv 2005.11401, 2020)
- OWASP, LLM08:2025 Vector and Embedding Weaknesses
- Loi n° 19-13 du 11 décembre 2019 régissant les activités d'hydrocarbures, Journal officiel n° 79 du 22 décembre 2019 (FAOLEX)
Frequently asked questions
Can an LLM search scanned well reports from the 1990s?
Only after OCR has turned the scans into text. Typed reports usually read well, but numbers, tables and handwritten notes need checking by hand on a sample. Keep a link from every extracted passage to the page image so engineers can verify depths and values in the original.
Do we need to fine-tune a model on drilling jargon?
Usually not as a first step. A glossary of abbreviations and activity codes, used to expand questions and to tag passages at indexing time, solves much of the jargon problem. Fine-tuning can be considered later if tests show the model still misreads your reports, but it does not give citations or keep up with new reports.
How do we handle wells whose names are written differently over the years?
Build a well master list with one identifier per wellbore and every alias found in the archive, such as names with slashes, underscores, prefixes or sidetrack suffixes. Map every report to that identifier when it is indexed, and filter searches on it.
Should NPT figures come from the LLM or from structured data?
From structured data. Durations, depths and counts should be computed by a database query over activity records, not added up by the language model from text. The model can choose and explain the query, then cite the reports behind the figures.
Can the system run without internet access?
Yes. OCR, embeddings, search indexes and the language model can all run on servers inside the company network, including on isolated networks. Updates then have to be brought in through a controlled process.
Who may use upstream well data in Algeria?
Law 19-13 on hydrocarbon activities states that upstream data are the exclusive property of the State and are managed by ALNAFT, which sets the conditions of access and distribution. Confirm with legal counsel, and with ALNAFT where relevant, what your company may do with the reports before indexing them or sharing them with a vendor.



