Salience Data

Data and Import

LinkedIn access is restricted, inconsistent, and changes without notice. This module exists so that no capability depends on a single access path, and so the system degrades to something useful rather than failing.

Design rule: every capability that wants LinkedIn data has a supported path, a user-assisted fallback, and an import-based fallback. The fallback is not a degraded mode — for most executive work, pasted text is entirely sufficient, and saying so avoids pushing the user toward tooling they do not need.

Retrieval tiers

Choose the highest tier available. Report which tier was used when it affects confidence.

Tier 0a — A corpus the user already keeps (best available, if it exists)

Many executives already have a career folder: resumes, press, awards, case studies, prior applications, writing samples. Pointing Salience at it beats every other intake path, because it is already organized, already dated, and already reconciled.

My career material is in <a folder> and my writing samples are in <another> — use those.

A source can be a directory, a single consolidated file, or a git repository, and there can be any number of them. Paths are personal and live in $SALIENCE_HOME/config.yaml; nothing in this plugin names a real location.

Survey first, then read only what the plan names. Never ingest a directory wholesale — a career folder with a code project inside it is common, and the file count will be dominated by node_modules. Report what was read, not just what was found.

Full method — configuration, folder-signal classification, exclusions, tiering by artifact type, provenance, and privacy: references/corpus-ingestion.md.

Tier 0 — User-provided (default, always available, zero setup)

The user pastes profile sections, a post body, or a job description; or supplies a document. No credentials, no third party, no terms-of-service exposure.

This is the default and it is not a compromise. A profile audit, a positioning exercise, a post draft, and a comment all work perfectly from pasted text. Do not recommend a paid data path for work that does not need one.

Tier 1 — LinkedIn data export (best for the user's own profile)

Settings → Data privacy → Get a copy of your data. Arrives as CSV archives covering profile, positions, education, skills, connections, and messages.

The most complete and most reliable source of the user's own history, and it is first-party. Prefer it for onboarding. Import only the files needed — do not ingest the messages archive by default.

Tier 2 — Official API via an MCP server (for publishing and own-post data)

Where a LinkedIn MCP server is configured with proper OAuth, use it for: publishing posts and comments, reading the user's own posts and engagement statistics, and managing pages the user administers.

Constraints that are real and worth stating: OAuth scopes gate everything, member and organization URNs must come from API calls rather than being constructed, and the official API does not support people search, reading third-party profiles, jobs, or messaging. A server claiming otherwise is using something other than the official API.

Tier 3 — Third-party retrieval services (optional, user-configured)

Commercial services can retrieve public post bodies, comment threads, and engagement lists without credentials. Useful for engagement analysis at volume.

Only ever used when the user has configured one deliberately. Salience never signs up for, selects, or recommends a specific vendor as a default, and it states the cost and the terms-of-service consideration when one is used.

Never

• Collecting, storing, or using a LinkedIn password, cookie, or session token
• Reusing a captured browser session
• Bypassing CAPTCHAs or bot detection
• Evading rate limits
• Automating actions through the logged-in interface as if human

These are refused regardless of who asks, how the request is framed, or what a configured tool offers. If a request requires one, say what is being refused and offer the highest available supported path instead.

Import and parsing

Résumé or CV

Extract roles, titles, organizations, dates, scope, achievements, education, credentials. Every number found becomes a ledger candidate with the document as source. Flag anything ambiguous rather than resolving it — a date range written "2021–2023" does not tell you the months.

LinkedIn export

Map the CSVs to the identity schema. Positions.csv and Profile.csv carry the spine. Cross-check against any résumé already imported and surface every disagreement.

Pasted profile text

Section boundaries are usually inferable from structure. Confirm the parse before writing to the ledger — a mis-parsed heading silently corrupts the record.

Bios, case studies, decks

Rich in verified metrics and usually the best source of proof points, because they were written for an audience that could check them. Extract numbers with their context intact.

Post and engagement data

Post bodies, comment threads, engagement rosters. Available at Tier 0 (paste), Tier 2 (own posts), or Tier 3.

Engagement rosters by paste

Who reacted to or commented on a post is real signal — someone who engaged is warmer than a cold contact — and there is no API for it. Every tool in the wild that offers it does so through a captured session, which Salience refuses. The paste path recovers the whole capability in about fifteen seconds and needs no credential:

Open the post's reactions or comments list on LinkedIn, select the list, copy, and paste it here. I will parse and rank it.

What arrives is messy — names, headlines, degree badges, and "Follow" buttons interleaved, often with duplicates where someone both reacted and commented. Parse leniently:

• One person per name-plus-headline pair; drop UI chrome, follower counts, and button labels
• Collapse duplicates, keeping the strongest engagement signal for that person
• A commenter outranks a reactor. Someone who commented and reacted outranks both
• Where the paste is ambiguous, say what did not parse rather than guessing at a name

Then rank against what is already known: existing relationship records first, then people whose role matches what the user is pursuing, then the rest. Hand the result to salience-engage as a worklist.

These are other people's names, and governance's third-party rules apply in full. Score the roster in the session; persist only the people the user actually acts on, with professional context only. A pasted roster of forty names is not a contact list to keep — retaining it whole would be exactly the cross-source aggregation about private individuals that governance forbids. Say so if the user asks to save it, and offer to keep the ones they contacted.

Normalization

Everything lands in the identity schema (config/identity.schema.json).

• Dates to YYYY-MM. Where only a year is known, record the year and mark the precision — do not invent a month.
• Organizations to a canonical name, with former names retained. A company that renamed is one employer, not two.
• Titles kept verbatim, with an industry-standard equivalent recorded alongside where the internal title was idiosyncratic. Never overwrite the real title.
• Metrics with unit, period, and scope preserved. "38%" is not a fact; "blended CAC down 38% over 11 months across three brands" is.
• Every imported fact carries its source document and tier. Import is stated unless the document is itself an artifact of record, in which case it is verified with the document cited.

Identity resolution

When resolving a person or company to a profile:

• Never assert a match on name alone. Common names produce wrong matches, and a wrong match in a relationship record or a prospect brief is worse than no match.
• Corroborate on at least two signals: employer, title, location, mutual connection, or a distinguishing detail.
• Report the match confidence. When ambiguous, present the candidates and ask.
• Do not compile personal information beyond professional context, and do not aggregate across sources to build a profile of a private individual.

Failure handling

When a path fails, say what happened, what it means, and what to do — never fail silently and never fall back to a prohibited method.

Couldn't retrieve that post automatically — no retrieval service is configured.

Paste the post text and I'll work from that. It costs you one copy-paste and there's no
difference in the result.

If a configured service errors or rate-limits, report it plainly, wait, or fall back to Tier 0. Never retry aggressively against a rate limit.

References

• references/corpus-ingestion.md — pointing Salience at a directory the user already keeps
• references/import-paths.md — per-source parsing detail, LinkedIn export file map
• references/tiers.md — capability-by-tier matrix and what each path can and cannot do