Skip to content

Archive intelligence

Years of mail accumulate facts: prices agreed, deadlines slipped, vendors picked. mxr ask and mxr decisions make that corpus answerable without pretending the model “remembers” anything. Retrieval runs on the local lexical and semantic indexes; every claim cites a message id; uncited LLM output is rejected before it reaches you.

Records answer first, mxr ask answers the rest

Section titled “Records answer first, mxr ask answers the rest”

Archive’s answer box (mxr records ask, / in Archive) looks for the answer in record fields first, with no model: “lisbon booking ref” returns the booking reference from the record. Only when no record matches every word does it fall back to mxr ask over all your mail, and it says so: “No record matches “boiler warranty”. Searching all mail instead.” Use mxr ask directly for questions records can’t answer, such as what someone decided.

Terminal window
mxr ask "what did Alice and I decide about pricing in Q2?"

What you get: a Markdown answer that names the decision, followed by Citations: with the cited message ids and short quotes. If the retrieved candidates don’t support an answer, the response is literally “not enough evidence” — no synthesized confidence, no invented dates.

Date and sender filters are enforced by the daemon, not the LLM. The model only ever sees messages the filter already accepted.

Terminal window
# Narrow to a person and a date range:
mxr ask "pricing decisions" \
--from alice@example.com \
--after 2026-04-01 --before 2026-06-30 \
--format json

What you get: a JSON object { text, citations, retrieval }. citations lists { message_id, thread_id, subject, date, quote }. retrieval shows the requested vs. executed mode (hybrid, lexical, semantic) and candidate_count — so degraded queries are visible, not silent.

Terminal window
# Hybrid (default): lexical BM25 + semantic, reciprocal-rank fused.
mxr ask "infra postmortems" --mode hybrid
# Lexical-only — useful when the semantic index is rebuilding,
# or when you want exact-term matches.
mxr ask "INC-1234 root cause" --mode lexical
# Semantic-only — broader recall on paraphrase-heavy questions.
mxr ask "what did we say about latency budgets" --mode semantic

When --mode semantic is requested but the semantic profile isn’t ready, the daemon falls back to lexical and reports the executed mode in the JSON. Scripts can branch on .retrieval.executed_mode.

The cited message ids are scriptable:

Terminal window
mxr ask "who owns the legal review for v2" --format json \
| jq -r '.citations[].message_id' \
| xargs -I{} mxr cat {} --view reader

What you get: every cited message rendered with reader mode, in order, so you can verify the synthesis yourself.

mxr ask requires the LLM for synthesis. With [llm] enabled = false, the command returns the top retrieved candidates and a synthesis unavailable notice — so you still get the search hits, you just don’t get an answer paragraph. Combine with mxr search if you need a pure retrieval surface.

Terminal window
# Equivalent retrieval surface without synthesis:
mxr search 'from:alice@example.com "pricing" newer_than:90d' --format json

Important decisions land in long threads. mxr decisions extracts them into a stable ledger you can query by topic, time, or stable id. Each row carries the threads and message ids that justified the extraction.

Terminal window
# List the most recent decisions.
mxr decisions --format json
# Filter by topic tag (extracted by the LLM during rebuild):
mxr decisions --topic pricing --since 180 --format json
# Show one row in full (every evidence message + participants):
mxr decisions show DECISION_ID --format json

What you get: rows { id, thread_id, decision, topic_tags, participants, evidence, decided_at, status }. id is hash(account, thread, normalized_decision, evidence_ids) — stable across rebuilds. decided_at is set only when the thread says when the decision was made (for example “on Friday”), resolved against the message that cites it; otherwise it is empty.

The log is populated incrementally on thread ingest. To re-extract over a window (e.g. after upgrading a model or after tuning the prompt):

Terminal window
# Dry-run-friendly: rebuild is idempotent on unchanged content.
mxr decisions rebuild --since 180
# Different window, JSON for scripting:
mxr decisions rebuild --since 365 --format json

What you get: a JSON envelope with processed_thread_count, extracted_count, and kept_count. Unchanged threads are skipped via source-hash comparison.

Terminal window
# For every pricing decision in the last quarter, open the thread it came from.
mxr decisions --topic pricing --since 90 --format json \
| jq -r '.[] | .thread_id' \
| xargs -I{} mxr thread {} --format json \
| jq -r '.subject'

What you get: subjects of every thread that produced a pricing decision in the window. Use --format ids instead of the jq pipe if you only want thread ids.

  • “What did we ship?” at quarter end: mxr decisions --since 90 --format json | jq -r '.[] | "\(.decided_at)\t\(.topic_tags | join(","))\t\(.decision)"' | sort — chronological list of decisions tagged by topic, every one cite-backed.
  • Onboarding a new teammate: mxr ask "how do we run production deploys?" --mode hybrid --limit 20 — synthesized answer plus the exact messages they should read.
  • Pre-meeting prep: mxr ask "open items with Bob" --from bob@example.com --after 2026-01-01 — one-paragraph state of the relationship before walking into a 1:1.
  • Audit after a postmortem: mxr decisions rebuild --since 30 && mxr decisions --topic incident --format json — re-extract after the postmortem thread closes, then list every decision the team made.
"Use `mxr ask "<question>" --format json` to answer the question below.
Only quote the model's `.text` field. Append every entry from
`.citations[]` verbatim. If `.retrieval.executed_mode` differs from
"hybrid", flag that to me. Question: {{question}}"
"For each topic in [hiring, pricing, infra], run `mxr decisions --topic
<t> --since 180 --format json` and produce a 5-bullet summary. Cite
the `evidence_msg_ids` for every bullet. Don't synthesize anything not
covered by an evidence id."