salesforce-documentation-context

Usage (Context7 MCP)

Install the Context7 MCP server into your AI agent: https://github.com/upstash/context7#installation

Then tell the agent to use this library (the llms.txt from this repo) directly via Context7. Example instruction:

- If a Context7 MCP server is available, call `query-docs` directly (skip `resolve-library-id`) with
  `context7CompatibleLibraryID: damecek/salesforce-documentation-context`, to get up to date information about target topic.

The live Context7 server (v3.2.2) renamed get-library-docs → query-docs. Always pin the library explicitly — see the Benchmark below for why unpinned resolution is unsafe for Salesforce.

Benchmark

We benchmarked this Context7 library against the official Salesforce Docs MCP across 11 real Salesforce questions (SOQL, Apex, Flow, LWC). Full report, methodology and raw answers: compare/REPORT.md (reproduce with python3 compare/compare.py --all).

Average score (relevance + correctness, 0–5):

Approach	Avg	Notes
A · Salesforce Docs MCP (official)	4.3	Broadest, freshest, best on procedures; raw JSON chunks, largest payloads
B · Context7 + this library (pinned)	3.0	Compact, deterministic, code-first, low token cost; limited to our corpus
C · Context7 without pinning the library	0.95	Resolver drifts to the wrong ecosystem (Go, Ruby, Flutter, WebVR…) — unusable

Conclusion — stay with Context7, don't migrate. The official Docs MCP wins overall, but it doesn't make this library obsolete: our pinned corpus is the low-token, deterministic, code-first option that stays 100% Salesforce-relevant and cites stable GitHub source links. The two are complements, not substitutes:

• Use the official Salesforce Docs MCP for broad platform / admin / how-to / procedural questions.
• Use Context7 pinned to /damecek/salesforce-documentation-context for compact, code-grounded answers from our curated corpus.
• Never use Context7 unpinned for Salesforce — approach C proves the resolver ignores platform context and picks the wrong library most of the time.

Migrating away from Context7 would trade a cheap, deterministic complement for nothing. The bigger win is closing corpus gaps the benchmark surfaced (LWC Developer Guide, LWR/LWS) by feeding them into src.txt.

Generate Documentation

One place where Salesforce documentation (PDFs on URLs) is downloaded, converted to Markdown, and written into documentation/. The project also generates llms.txt (per https://llmstxt.org/) with links to the GitHub Raw versions of those Markdown files.

Requirements

• Python 3.10+ (python3 --version)
• uv (uv --version)

Fastest way to create or re-generate Markdown documentation files (documentation/) and update llms.txt:

1. Update src.txt (your list of PDF sources).
2. Install dependencies:

uv sync

3. Generate / refresh Markdown:

FORCE=1 PDF_IGNORE_IMAGES=1 PDF_FORCE_TEXT=1 uv run update

Output:

• Markdown files in documentation/ (one or more per entry in src.txt; large outputs are split).
• Download cache in .cache/ (ETag / Last-Modified when available).
• Updated llms.txt with GitHub Raw links.

Refresh only one specific link

Do not use FORCE=1 here (it clears the whole documentation/ folder).

1. Prepare a temporary source file with just the one entry:

cat > /tmp/one.txt <<'EOF'
https://example.com/doc.pdf | doc.md
EOF

2. (Optional but recommended) Clear the cache just for that URL so it always regenerates:

URL="https://example.com/doc.pdf"
KEY=$(python - <<'PY'
import hashlib, os
url = os.environ["URL"]
print(hashlib.sha256(url.encode("utf-8")).hexdigest()[:24])
PY
)
rm -f ".cache/${KEY}.pdf" ".cache/${KEY}.json"

3. Run the updater against only that temporary file:

uv run update --src /tmp/one.txt

Conversion notes:

• Uses pymupdf4llm.to_markdown() (PyMuPDF4LLM) for Markdown output.
• Header detection defaults to auto (TOC-based if available, otherwise font-size heuristics).

src.txt format

Supported variants (1 line = 1 document):

• https://example.com/doc.pdf
• https://example.com/doc.pdf | sfdx_dev.md
• Document title | https://example.com/doc.pdf
• Document title | https://example.com/doc.pdf | sfdx_dev.md

Lines starting with # are comments.

Configuration (optional)

The updater supports these env vars:

• SRC_FILE (default src.txt)
• DOCS_DIR (default documentation)
• CACHE_DIR (default .cache)
• CONCURRENCY (default 4)
• FORCE (set to 1/true to regenerate Markdown even if the PDF content is unchanged)
• HTTP_TIMEOUT (default 60, seconds)
• MAX_MD_BYTES (default 899999; split markdown outputs so each file stays under this size)
• GITHUB_RAW_BASE (e.g. https://raw.githubusercontent.com/OWNER/REPO/main)
• GITHUB_RAW_BRANCH (override branch used for raw links)

PyMuPDF4LLM conversion tuning:

• PDF_HDR_MODE (auto | toc | identify | none) (default auto)
• PDF_HDR_BODY_LIMIT (default 12)
• PDF_HDR_MAX_LEVELS (default 4)
• PDF_HDR_SAMPLE_PAGES (default 10)
• PDF_FORCE_TEXT (1/true to extract text even over images)
• PDF_IGNORE_IMAGES (1/true)
• PDF_IGNORE_GRAPHICS (1/true)
• PDF_TABLE_STRATEGY (e.g. lines_strict, lines, text, explicit; empty disables)