deep-research

Name: deep-research
Author: zilliztech

111stars13forks

How to install deep-research?

zilliztech/mfs/deep-research

$npx -y skills add zilliztech/mfs --skill deep-research

Installs into the current project.

Summary

Answer a broad, open-ended, or multi-part question by iteratively searching MFS-indexed sources, following up on gaps, and synthesizing a cited report — the same "reason and search over private data" job zilliztech/deep-searcher does, but as a skill on top of mfs-find instead o

Files · 1

View on GitHub

SKILL.md

1	# Deep research over MFS-indexed sources
2
3	## 1. What this is (and what it replaces)
4
5	[deep-searcher](https://github.com/zilliztech/deep-searcher) is an
6	open-source framework that reasons over private data: it decomposes a
7	question, iteratively searches a vector database, evaluates whether the
8	evidence is sufficient, and synthesizes a cited report. Built before
9	agentic coding tools existed, it had to hand-roll every piece itself —
10	document loaders, a multi-provider LLM/embedding/vector-DB matrix, and a
11	custom iterative-retrieval orchestration loop in Python.
12
13	None of that orchestration is needed anymore. MFS already does ingestion +
14	hybrid search over many source types, and an agent's own reasoning loop
15	already does "search, judge, follow up, repeat" natively once it has a
16	search tool. This skill is that missing piece: not new retrieval code, just
17	the strategy for running deep-searcher's decompose → search → evaluate
18	→ synthesize loop through `mfs search` / `mfs cat`.
19
20	This skill assumes `mfs-find` for the actual command mechanics (search
21	modes, locators, `--peek`/`--skim`, index-status diagnosis). Read that
22	skill for those details — this one only adds the multi-round strategy on
23	top.
24
25	## 2. Precondition: sources must be indexed
26
27	Same as `mfs-find`: `mfs status` / `mfs connector inspect <uri>` first. If
28	nothing relevant is indexed yet, redirect to `mfs-ingest` — don't run a
29	research loop against an empty index.
30
31	## 3. The loop
32
33	```
34	decompose search rounds evaluate synthesize
35	┌───────────┐ ┌───────────────────────┐ ┌──────────────────┐ ┌───────────┐
36	│ 2-4 angles│ → │ mfs search per angle, │ → │ enough coverage? │ → │ cited │
37	│ on the Q │ │ semantic + keyword │ │ gaps → new angles│ │ report │
38	└───────────┘ └───────────────────────┘ └──────┬───────────┘ └───────────┘
39	▲ │ not enough
40	└────────────────────────────┘ (max ~4 rounds)
41	```
42
43	1. Decompose. Break the question into 2-4 concrete angles before
44	searching anything. "Write a report on our rate-limiting story" becomes:
45	current implementation, past incidents/bugs, *design
46	discussion/rationale, config knobs*. A single search for the raw
47	question under-recalls on anything but the narrowest asks.
48
49	2. Search each angle, scope per `mfs-find` §6-7 (hybrid by default,
50	`--all` for genuinely cross-source asks, scoped to 2-3 likely sources
51	otherwise):
52	```bash
53	mfs search "<angle 1>" <scope> --top-k 15
54	mfs search "<angle 2>" <scope> --top-k 15
55	...
56	```
57	Track distinct objects found (dedupe by `source`), not raw hit
58	count — five chunks from one file is one citation, not five.
59
60	3. Evaluate coverage before reading everything. For each angle: is
61	there at least one strong hit? For the question as a whole: do the
62	found objects, read together, actually answer it, or do they only
63	establish that the topic exists? Common gaps: an angle returned nothing
64	(rephrase it, don't drop it silently), all hits are from one source
65	type when the question implies more (e.g. only code, no design docs),
66	or a hit references something ("see the migration doc") not yet found.
67
68	4. Follow up, don't restart. Gaps become 1-3 new targeted searches —
69	reuse the vocabulary actually found in round 1 (a real error code, a
70	real doc title) instead of guessing more synonyms of the original
71	question. This is the step deep-searcher's LLM-driven query refinement
72	automated; here it's just another `mfs search` call informed by what
73	came back.
74
75	5. Stop condition. Whichever comes first:
76	- a round adds no new distinct objects (saturation), or
77	- every decomposed angle has a strong hit and no unresolved reference
78	remains, or
79	- ~4 rounds (glance at whether the effort is still paying off,
80	don't hard-stop exactly at 4 if one more obvious query would close a
81	re

People who liked this also viewed

6 picks

Type

deep-research

111stars13forks

View on GitHub

How to install deep-research?

zilliztech/mfs/deep-research

$npx -y skills add zilliztech/mfs --skill deep-research

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Deep research over MFS-indexed sources
2
3	## 1. What this is (and what it replaces)
4
5	[deep-searcher](https://github.com/zilliztech/deep-searcher) is an
6	open-source framework that reasons over private data: it decomposes a
7	question, iteratively searches a vector database, evaluates whether the
8	evidence is sufficient, and synthesizes a cited report. Built before
9	agentic coding tools existed, it had to hand-roll every piece itself —
10	document loaders, a multi-provider LLM/embedding/vector-DB matrix, and a
11	custom iterative-retrieval orchestration loop in Python.
12
13	None of that orchestration is needed anymore. MFS already does ingestion +
14	hybrid search over many source types, and an agent's own reasoning loop
15	already does "search, judge, follow up, repeat" natively once it has a
16	search tool. This skill is that missing piece: not new retrieval code, just
17	the strategy for running deep-searcher's decompose → search → evaluate
18	→ synthesize loop through `mfs search` / `mfs cat`.
19
20	This skill assumes `mfs-find` for the actual command mechanics (search
21	modes, locators, `--peek`/`--skim`, index-status diagnosis). Read that
22	skill for those details — this one only adds the multi-round strategy on
23	top.
24
25	## 2. Precondition: sources must be indexed
26
27	Same as `mfs-find`: `mfs status` / `mfs connector inspect <uri>` first. If
28	nothing relevant is indexed yet, redirect to `mfs-ingest` — don't run a
29	research loop against an empty index.
30
31	## 3. The loop
32
33	```
34	decompose search rounds evaluate synthesize
35	┌───────────┐ ┌───────────────────────┐ ┌──────────────────┐ ┌───────────┐
36	│ 2-4 angles│ → │ mfs search per angle, │ → │ enough coverage? │ → │ cited │
37	│ on the Q │ │ semantic + keyword │ │ gaps → new angles│ │ report │
38	└───────────┘ └───────────────────────┘ └──────┬───────────┘ └───────────┘
39	▲ │ not enough
40	└────────────────────────────┘ (max ~4 rounds)
41	```
42
43	1. Decompose. Break the question into 2-4 concrete angles before
44	searching anything. "Write a report on our rate-limiting story" becomes:
45	current implementation, past incidents/bugs, *design
46	discussion/rationale, config knobs*. A single search for the raw
47	question under-recalls on anything but the narrowest asks.
48
49	2. Search each angle, scope per `mfs-find` §6-7 (hybrid by default,
50	`--all` for genuinely cross-source asks, scoped to 2-3 likely sources
51	otherwise):
52	```bash
53	mfs search "<angle 1>" <scope> --top-k 15
54	mfs search "<angle 2>" <scope> --top-k 15
55	...
56	```
57	Track distinct objects found (dedupe by `source`), not raw hit
58	count — five chunks from one file is one citation, not five.
59
60	3. Evaluate coverage before reading everything. For each angle: is
61	there at least one strong hit? For the question as a whole: do the
62	found objects, read together, actually answer it, or do they only
63	establish that the topic exists? Common gaps: an angle returned nothing
64	(rephrase it, don't drop it silently), all hits are from one source
65	type when the question implies more (e.g. only code, no design docs),
66	or a hit references something ("see the migration doc") not yet found.
67
68	4. Follow up, don't restart. Gaps become 1-3 new targeted searches —
69	reuse the vocabulary actually found in round 1 (a real error code, a
70	real doc title) instead of guessing more synonyms of the original
71	question. This is the step deep-searcher's LLM-driven query refinement
72	automated; here it's just another `mfs search` call informed by what
73	came back.
74
75	5. Stop condition. Whichever comes first:
76	- a round adds no new distinct objects (saturation), or
77	- every decomposed angle has a strong hit and no unresolved reference
78	remains, or
79	- ~4 rounds (glance at whether the effort is still paying off,
80	don't hard-stop exactly at 4 if one more obvious query would close a
81	re

People who liked this also viewed

6 picks

Type