bulk-operations

Bulk Operations

Batch ingestion of Plaud recordings into the memory search index (Cognee with Qdrant vectors + Neo4j graph). Handles transcription, chunking, embedding, and graph construction at scale.

SAFETY PROTOCOL

1. ALWAYS run dry_run=true first for any bulk operation
2. Review the cost estimate with the user before proceeding
3. Prefer source="cache" (free) over source="all" (costs ElevenLabs credits)
4. Use since to limit scope when possible
5. Track progress with job_id returned from initial call

Tools

memory_ingest

Main bulk ingestion tool. Transcribes, chunks, embeds, and indexes recordings into the memory search pipeline.

Parameter	Type	Default	Description
source	"all" | "cache"	"all"	"all": transcribe missing (COSTS CREDITS). "cache": cached only (FREE)
mode	"once" | "auto"	"once"	"once": single run. "auto": background poller
action	string	null	Required when mode="auto": "start", "stop", or "status"
file_ref	string	null	Ingest a single recording only
since	string	null	Only recordings after date ("7d", "30d", "2026-01-01")
language	string	"eng"	ElevenLabs 3-letter language code
dry_run	bool	false	Show cost estimate without processing
job_id	string	null	Poll status of a running job

Usage Patterns

memory_ingest(dry_run=true)                    # Cost estimate for all recordings
memory_ingest(source="cache")                  # Index only cached transcripts (FREE)
memory_ingest(source="all")                    # Transcribe + index everything (COSTS CREDITS)
memory_ingest(file_ref="abc123")               # Ingest a single recording
memory_ingest(since="7d")                      # Only recent recordings
memory_ingest(since="7d", dry_run=true)        # Cost estimate for recent only
memory_ingest(job_id="abc-123")                # Poll job progress
memory_ingest(mode="auto", action="start")     # Start background auto-poller
memory_ingest(mode="auto", action="status")    # Check auto-poller status
memory_ingest(mode="auto", action="stop")      # Stop auto-poller

Workflows

"Index all my recordings"

1. memory_ingest(dry_run=true) -- review cost estimate
2. Present estimate to user, get confirmation
3. If user approves: memory_ingest(source="all")
4. Track with returned job_id: memory_ingest(job_id="<id>")

"Index only what's already transcribed (free)"

1. memory_ingest(source="cache") -- no ElevenLabs cost
2. Track with job_id

"Index just the last week"

1. memory_ingest(since="7d", dry_run=true) -- estimate
2. memory_ingest(since="7d") -- execute

"Index a single recording"

1. memory_ingest(file_ref="<id>") -- transcribes + indexes one recording

"Set up continuous ingestion"

1. Explain: auto-poller will continuously monitor for new recordings and transcribe them (costs credits for each new recording)
2. Get user confirmation
3. memory_ingest(mode="auto", action="start")
4. Check status: memory_ingest(mode="auto", action="status")
5. Stop when done: memory_ingest(mode="auto", action="stop")

"Check job progress"

1. memory_ingest(job_id="<id>") -- returns progress, count, errors

source="all" vs source="cache"

	source="all"	source="cache"
Cost	ElevenLabs API credits per uncached recording	Free
Coverage	All recordings (transcribes missing ones)	Only already-transcribed recordings
Speed	Slower (transcription per recording)	Fast (skip transcription step)
Use when	User wants everything indexed	User wants to avoid cost

Anti-Patterns (NEVER DO)

• Never run memory_ingest(source="all") without dry_run=true first -- always estimate cost
• Never start auto-poller without explaining that it will continuously transcribe new recordings at cost
• Never ignore job_id -- always poll for completion to confirm success
• Never run bulk ingestion without user confirmation -- cost can be significant
• Never assume ingestion succeeded -- check job_id status and verify with memory_search