[bernardladenthin]

docs/local-inference-presentation.md now tells the map of the source: why an agent needs it, what it looks like for any kind of source, whether it leads to the right file, and how to run it. The numbers come from the model card and the re-measurements of 2026-11-01/02.

Main part (13 slides)

1. Agents need a map of the source first (new)
2. srcmorph: a navigable map: the benchmark corpus as the example tree, leads, exact facts, incremental updates (reworked)
3. One entry of the map: a real summary (870 KB XML catalog, GPT-6 Luna) with header, facts line and coverage line (new)
4. Any kind of source: prompt sections and fact counters for Java, SQL, shell, Perl, XML and the fallback; the language results of finding 7 (new)
5. Does the map lead to the right file? The search measurement: 86–98 % first with the summaries, 74 % with paths, 72 % with the whole source at about 4× the tokens (new)
6. How compact the map is (unchanged)
7. Large files: map-reduce over every chunk, the facts and coverage lines, the planted middle names (new)
8. Choosing a model: hosted from profile v3, local from the card (reworked; the old slide said gpt-oss-20b does not fit 8 GB, but the card has it locally at 95.8 %, 33.1 s per file)
9. Running it: one configuration file, plan/check, incremental, run record, providers (new)
10. The foundation: java-llama.cpp (the old slides 1 and 2 merged)
11. Outlook (unchanged)
12. Next: what to measure (new)
13. Top 5: what the map brings to a large project, each point with its measured number: a new chat starts with the whole tree in view (project entry 3.06 KB, 1.3 % of the source); questions in plain words find the right file first (86–98 % against 74 / 72 %; BM25 gains nothing); about a quarter of the input tokens (≈ 16,000 against ≈ 65,000); large files no window holds (GPT-6 Luna 99.0 %, every planted middle name); 0.002–0.003 USD per index run, incremental, local without an API fee. The footer says it is measured on 24 files and that a large project is not yet measured (new)

Appendix (slides 14–17): native binaries and the three charts. Their texts now say 12 of the card's 35 runs (they said 20), and the heatmap is labeled as the charts' ten models rather than "Top 10". The chart files are unchanged: re-rendering them would also move the model card's charts, whose published numbers stay.

PDF: docs/local-inference-presentation.pdf (17 pages, 16:9) is committed next to the deck and written by docs/render-presentation.py: one HTML page with a 1280×720 section per slide and the front matter's style, printed by headless Chrome or Edge, so neither Node nor the Marp CLI is needed. Standard library only; it finds the browser on Windows, macOS and Linux and fails when the page count differs from the slide count. Render again after every change of the deck or a deck chart: python docs/render-presentation.py.

Checked: every page of the PDF inspected; the script reproduces the committed PDF (identical text); REUSE compliant.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DuAkLtHrCTx81vP5Mt4jad
https://claude.ai/code/session_01S8cAMa6XrThLWn4hooEtrA

Summary by CodeRabbit

• Documentation
  • Added a license notice to the local inference presentation PDF.
  • Added a way to generate a PDF version of the presentation using an available browser, with options to choose the browser and save an HTML preview.
  • PDF generation checks that the resulting document contains the expected number of slides and reports the slide count when successful.

[coderabbitai]

Caution

Review failed

The pull request is closed.

Walkthrough

A new script converts the local inference presentation deck into a PDF using a headless browser. The change also adds copyright and Apache-2.0 SPDX metadata for the PDF.

Changes: Presentation PDF

Layer / File(s)	Summary
Browser selection and HTML preparation docs/render-presentation.py	The script locates an available browser and converts the deck into styled HTML sections. It counts the sections as slides.
PDF rendering and validation docs/render-presentation.py, docs/local-inference-presentation.pdf.license	The command-line workflow prints the HTML to PDF and checks that the PDF page count matches the slide count. The license file adds a 2026 copyright attribution and an Apache-2.0 identifier.

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Other

Sequence Diagram(s)

sequenceDiagram
  participant Renderer as render-presentation.py
  participant Deck as local-inference-presentation.md
  participant Browser as Headless browser
  participant PDF as PDF output
  Renderer->>Deck: Read deck text
  Renderer->>Renderer: Build HTML and count slides
  Renderer->>Browser: Print HTML to PDF
  Browser->>PDF: Write PDF
  Renderer->>PDF: Check PDF page count

Loading

Suggested reviewers: claude

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud